xref: /illumos-gate/usr/src/uts/common/fs/zfs/metaslab.c (revision 084fd14f7c3336eb67ee283cabad2da8998b00d6)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
24  * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25  * Copyright (c) 2014 Integros [integros.com]
26  * Copyright (c) 2017, Intel Corporation.
27  */
28 
29 #include <sys/zfs_context.h>
30 #include <sys/dmu.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/space_map.h>
33 #include <sys/metaslab_impl.h>
34 #include <sys/vdev_impl.h>
35 #include <sys/zio.h>
36 #include <sys/spa_impl.h>
37 #include <sys/zfeature.h>
38 #include <sys/vdev_indirect_mapping.h>
39 #include <sys/zap.h>
40 
41 #define	GANG_ALLOCATION(flags) \
42 	((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER))
43 
44 uint64_t metaslab_aliquot = 512ULL << 10;
45 uint64_t metaslab_force_ganging = SPA_MAXBLOCKSIZE + 1;	/* force gang blocks */
46 
47 /*
48  * Since we can touch multiple metaslabs (and their respective space maps)
49  * with each transaction group, we benefit from having a smaller space map
50  * block size since it allows us to issue more I/O operations scattered
51  * around the disk.
52  */
53 int zfs_metaslab_sm_blksz = (1 << 12);
54 
55 /*
56  * The in-core space map representation is more compact than its on-disk form.
57  * The zfs_condense_pct determines how much more compact the in-core
58  * space map representation must be before we compact it on-disk.
59  * Values should be greater than or equal to 100.
60  */
61 int zfs_condense_pct = 200;
62 
63 /*
64  * Condensing a metaslab is not guaranteed to actually reduce the amount of
65  * space used on disk. In particular, a space map uses data in increments of
66  * MAX(1 << ashift, space_map_blksize), so a metaslab might use the
67  * same number of blocks after condensing. Since the goal of condensing is to
68  * reduce the number of IOPs required to read the space map, we only want to
69  * condense when we can be sure we will reduce the number of blocks used by the
70  * space map. Unfortunately, we cannot precisely compute whether or not this is
71  * the case in metaslab_should_condense since we are holding ms_lock. Instead,
72  * we apply the following heuristic: do not condense a spacemap unless the
73  * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
74  * blocks.
75  */
76 int zfs_metaslab_condense_block_threshold = 4;
77 
78 /*
79  * The zfs_mg_noalloc_threshold defines which metaslab groups should
80  * be eligible for allocation. The value is defined as a percentage of
81  * free space. Metaslab groups that have more free space than
82  * zfs_mg_noalloc_threshold are always eligible for allocations. Once
83  * a metaslab group's free space is less than or equal to the
84  * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
85  * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
86  * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
87  * groups are allowed to accept allocations. Gang blocks are always
88  * eligible to allocate on any metaslab group. The default value of 0 means
89  * no metaslab group will be excluded based on this criterion.
90  */
91 int zfs_mg_noalloc_threshold = 0;
92 
93 /*
94  * Metaslab groups are considered eligible for allocations if their
95  * fragmenation metric (measured as a percentage) is less than or equal to
96  * zfs_mg_fragmentation_threshold. If a metaslab group exceeds this threshold
97  * then it will be skipped unless all metaslab groups within the metaslab
98  * class have also crossed this threshold.
99  */
100 int zfs_mg_fragmentation_threshold = 85;
101 
102 /*
103  * Allow metaslabs to keep their active state as long as their fragmentation
104  * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
105  * active metaslab that exceeds this threshold will no longer keep its active
106  * status allowing better metaslabs to be selected.
107  */
108 int zfs_metaslab_fragmentation_threshold = 70;
109 
110 /*
111  * When set will load all metaslabs when pool is first opened.
112  */
113 int metaslab_debug_load = 0;
114 
115 /*
116  * When set will prevent metaslabs from being unloaded.
117  */
118 int metaslab_debug_unload = 0;
119 
120 /*
121  * Minimum size which forces the dynamic allocator to change
122  * it's allocation strategy.  Once the space map cannot satisfy
123  * an allocation of this size then it switches to using more
124  * aggressive strategy (i.e search by size rather than offset).
125  */
126 uint64_t metaslab_df_alloc_threshold = SPA_OLD_MAXBLOCKSIZE;
127 
128 /*
129  * The minimum free space, in percent, which must be available
130  * in a space map to continue allocations in a first-fit fashion.
131  * Once the space map's free space drops below this level we dynamically
132  * switch to using best-fit allocations.
133  */
134 int metaslab_df_free_pct = 4;
135 
136 /*
137  * A metaslab is considered "free" if it contains a contiguous
138  * segment which is greater than metaslab_min_alloc_size.
139  */
140 uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
141 
142 /*
143  * Percentage of all cpus that can be used by the metaslab taskq.
144  */
145 int metaslab_load_pct = 50;
146 
147 /*
148  * Determines how many txgs a metaslab may remain loaded without having any
149  * allocations from it. As long as a metaslab continues to be used we will
150  * keep it loaded.
151  */
152 int metaslab_unload_delay = TXG_SIZE * 2;
153 
154 /*
155  * Max number of metaslabs per group to preload.
156  */
157 int metaslab_preload_limit = SPA_DVAS_PER_BP;
158 
159 /*
160  * Enable/disable preloading of metaslab.
161  */
162 boolean_t metaslab_preload_enabled = B_TRUE;
163 
164 /*
165  * Enable/disable fragmentation weighting on metaslabs.
166  */
167 boolean_t metaslab_fragmentation_factor_enabled = B_TRUE;
168 
169 /*
170  * Enable/disable lba weighting (i.e. outer tracks are given preference).
171  */
172 boolean_t metaslab_lba_weighting_enabled = B_TRUE;
173 
174 /*
175  * Enable/disable metaslab group biasing.
176  */
177 boolean_t metaslab_bias_enabled = B_TRUE;
178 
179 /*
180  * Enable/disable remapping of indirect DVAs to their concrete vdevs.
181  */
182 boolean_t zfs_remap_blkptr_enable = B_TRUE;
183 
184 /*
185  * Enable/disable segment-based metaslab selection.
186  */
187 boolean_t zfs_metaslab_segment_weight_enabled = B_TRUE;
188 
189 /*
190  * When using segment-based metaslab selection, we will continue
191  * allocating from the active metaslab until we have exhausted
192  * zfs_metaslab_switch_threshold of its buckets.
193  */
194 int zfs_metaslab_switch_threshold = 2;
195 
196 /*
197  * Internal switch to enable/disable the metaslab allocation tracing
198  * facility.
199  */
200 boolean_t metaslab_trace_enabled = B_TRUE;
201 
202 /*
203  * Maximum entries that the metaslab allocation tracing facility will keep
204  * in a given list when running in non-debug mode. We limit the number
205  * of entries in non-debug mode to prevent us from using up too much memory.
206  * The limit should be sufficiently large that we don't expect any allocation
207  * to every exceed this value. In debug mode, the system will panic if this
208  * limit is ever reached allowing for further investigation.
209  */
210 uint64_t metaslab_trace_max_entries = 5000;
211 
212 /*
213  * Maximum number of metaslabs per group that can be disabled
214  * simultaneously.
215  */
216 int max_disabled_ms = 3;
217 
218 static uint64_t metaslab_weight(metaslab_t *);
219 static void metaslab_set_fragmentation(metaslab_t *);
220 static void metaslab_free_impl(vdev_t *, uint64_t, uint64_t, boolean_t);
221 static void metaslab_check_free_impl(vdev_t *, uint64_t, uint64_t);
222 static void metaslab_passivate(metaslab_t *msp, uint64_t weight);
223 static uint64_t metaslab_weight_from_range_tree(metaslab_t *msp);
224 
225 kmem_cache_t *metaslab_alloc_trace_cache;
226 
227 /*
228  * ==========================================================================
229  * Metaslab classes
230  * ==========================================================================
231  */
232 metaslab_class_t *
233 metaslab_class_create(spa_t *spa, metaslab_ops_t *ops)
234 {
235 	metaslab_class_t *mc;
236 
237 	mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
238 
239 	mc->mc_spa = spa;
240 	mc->mc_rotor = NULL;
241 	mc->mc_ops = ops;
242 	mutex_init(&mc->mc_lock, NULL, MUTEX_DEFAULT, NULL);
243 	mc->mc_alloc_slots = kmem_zalloc(spa->spa_alloc_count *
244 	    sizeof (zfs_refcount_t), KM_SLEEP);
245 	mc->mc_alloc_max_slots = kmem_zalloc(spa->spa_alloc_count *
246 	    sizeof (uint64_t), KM_SLEEP);
247 	for (int i = 0; i < spa->spa_alloc_count; i++)
248 		zfs_refcount_create_tracked(&mc->mc_alloc_slots[i]);
249 
250 	return (mc);
251 }
252 
253 void
254 metaslab_class_destroy(metaslab_class_t *mc)
255 {
256 	ASSERT(mc->mc_rotor == NULL);
257 	ASSERT(mc->mc_alloc == 0);
258 	ASSERT(mc->mc_deferred == 0);
259 	ASSERT(mc->mc_space == 0);
260 	ASSERT(mc->mc_dspace == 0);
261 
262 	for (int i = 0; i < mc->mc_spa->spa_alloc_count; i++)
263 		zfs_refcount_destroy(&mc->mc_alloc_slots[i]);
264 	kmem_free(mc->mc_alloc_slots, mc->mc_spa->spa_alloc_count *
265 	    sizeof (zfs_refcount_t));
266 	kmem_free(mc->mc_alloc_max_slots, mc->mc_spa->spa_alloc_count *
267 	    sizeof (uint64_t));
268 	mutex_destroy(&mc->mc_lock);
269 	kmem_free(mc, sizeof (metaslab_class_t));
270 }
271 
272 int
273 metaslab_class_validate(metaslab_class_t *mc)
274 {
275 	metaslab_group_t *mg;
276 	vdev_t *vd;
277 
278 	/*
279 	 * Must hold one of the spa_config locks.
280 	 */
281 	ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
282 	    spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
283 
284 	if ((mg = mc->mc_rotor) == NULL)
285 		return (0);
286 
287 	do {
288 		vd = mg->mg_vd;
289 		ASSERT(vd->vdev_mg != NULL);
290 		ASSERT3P(vd->vdev_top, ==, vd);
291 		ASSERT3P(mg->mg_class, ==, mc);
292 		ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
293 	} while ((mg = mg->mg_next) != mc->mc_rotor);
294 
295 	return (0);
296 }
297 
298 static void
299 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
300     int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
301 {
302 	atomic_add_64(&mc->mc_alloc, alloc_delta);
303 	atomic_add_64(&mc->mc_deferred, defer_delta);
304 	atomic_add_64(&mc->mc_space, space_delta);
305 	atomic_add_64(&mc->mc_dspace, dspace_delta);
306 }
307 
308 uint64_t
309 metaslab_class_get_alloc(metaslab_class_t *mc)
310 {
311 	return (mc->mc_alloc);
312 }
313 
314 uint64_t
315 metaslab_class_get_deferred(metaslab_class_t *mc)
316 {
317 	return (mc->mc_deferred);
318 }
319 
320 uint64_t
321 metaslab_class_get_space(metaslab_class_t *mc)
322 {
323 	return (mc->mc_space);
324 }
325 
326 uint64_t
327 metaslab_class_get_dspace(metaslab_class_t *mc)
328 {
329 	return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
330 }
331 
332 void
333 metaslab_class_histogram_verify(metaslab_class_t *mc)
334 {
335 	spa_t *spa = mc->mc_spa;
336 	vdev_t *rvd = spa->spa_root_vdev;
337 	uint64_t *mc_hist;
338 	int i;
339 
340 	if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
341 		return;
342 
343 	mc_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
344 	    KM_SLEEP);
345 
346 	for (int c = 0; c < rvd->vdev_children; c++) {
347 		vdev_t *tvd = rvd->vdev_child[c];
348 		metaslab_group_t *mg = tvd->vdev_mg;
349 
350 		/*
351 		 * Skip any holes, uninitialized top-levels, or
352 		 * vdevs that are not in this metalab class.
353 		 */
354 		if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 ||
355 		    mg->mg_class != mc) {
356 			continue;
357 		}
358 
359 		for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
360 			mc_hist[i] += mg->mg_histogram[i];
361 	}
362 
363 	for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
364 		VERIFY3U(mc_hist[i], ==, mc->mc_histogram[i]);
365 
366 	kmem_free(mc_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
367 }
368 
369 /*
370  * Calculate the metaslab class's fragmentation metric. The metric
371  * is weighted based on the space contribution of each metaslab group.
372  * The return value will be a number between 0 and 100 (inclusive), or
373  * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
374  * zfs_frag_table for more information about the metric.
375  */
376 uint64_t
377 metaslab_class_fragmentation(metaslab_class_t *mc)
378 {
379 	vdev_t *rvd = mc->mc_spa->spa_root_vdev;
380 	uint64_t fragmentation = 0;
381 
382 	spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
383 
384 	for (int c = 0; c < rvd->vdev_children; c++) {
385 		vdev_t *tvd = rvd->vdev_child[c];
386 		metaslab_group_t *mg = tvd->vdev_mg;
387 
388 		/*
389 		 * Skip any holes, uninitialized top-levels,
390 		 * or vdevs that are not in this metalab class.
391 		 */
392 		if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 ||
393 		    mg->mg_class != mc) {
394 			continue;
395 		}
396 
397 		/*
398 		 * If a metaslab group does not contain a fragmentation
399 		 * metric then just bail out.
400 		 */
401 		if (mg->mg_fragmentation == ZFS_FRAG_INVALID) {
402 			spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
403 			return (ZFS_FRAG_INVALID);
404 		}
405 
406 		/*
407 		 * Determine how much this metaslab_group is contributing
408 		 * to the overall pool fragmentation metric.
409 		 */
410 		fragmentation += mg->mg_fragmentation *
411 		    metaslab_group_get_space(mg);
412 	}
413 	fragmentation /= metaslab_class_get_space(mc);
414 
415 	ASSERT3U(fragmentation, <=, 100);
416 	spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
417 	return (fragmentation);
418 }
419 
420 /*
421  * Calculate the amount of expandable space that is available in
422  * this metaslab class. If a device is expanded then its expandable
423  * space will be the amount of allocatable space that is currently not
424  * part of this metaslab class.
425  */
426 uint64_t
427 metaslab_class_expandable_space(metaslab_class_t *mc)
428 {
429 	vdev_t *rvd = mc->mc_spa->spa_root_vdev;
430 	uint64_t space = 0;
431 
432 	spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
433 	for (int c = 0; c < rvd->vdev_children; c++) {
434 		uint64_t tspace;
435 		vdev_t *tvd = rvd->vdev_child[c];
436 		metaslab_group_t *mg = tvd->vdev_mg;
437 
438 		if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 ||
439 		    mg->mg_class != mc) {
440 			continue;
441 		}
442 
443 		/*
444 		 * Calculate if we have enough space to add additional
445 		 * metaslabs. We report the expandable space in terms
446 		 * of the metaslab size since that's the unit of expansion.
447 		 * Adjust by efi system partition size.
448 		 */
449 		tspace = tvd->vdev_max_asize - tvd->vdev_asize;
450 		if (tspace > mc->mc_spa->spa_bootsize) {
451 			tspace -= mc->mc_spa->spa_bootsize;
452 		}
453 		space += P2ALIGN(tspace, 1ULL << tvd->vdev_ms_shift);
454 	}
455 	spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
456 	return (space);
457 }
458 
459 static int
460 metaslab_compare(const void *x1, const void *x2)
461 {
462 	const metaslab_t *m1 = (const metaslab_t *)x1;
463 	const metaslab_t *m2 = (const metaslab_t *)x2;
464 
465 	int sort1 = 0;
466 	int sort2 = 0;
467 	if (m1->ms_allocator != -1 && m1->ms_primary)
468 		sort1 = 1;
469 	else if (m1->ms_allocator != -1 && !m1->ms_primary)
470 		sort1 = 2;
471 	if (m2->ms_allocator != -1 && m2->ms_primary)
472 		sort2 = 1;
473 	else if (m2->ms_allocator != -1 && !m2->ms_primary)
474 		sort2 = 2;
475 
476 	/*
477 	 * Sort inactive metaslabs first, then primaries, then secondaries. When
478 	 * selecting a metaslab to allocate from, an allocator first tries its
479 	 * primary, then secondary active metaslab. If it doesn't have active
480 	 * metaslabs, or can't allocate from them, it searches for an inactive
481 	 * metaslab to activate. If it can't find a suitable one, it will steal
482 	 * a primary or secondary metaslab from another allocator.
483 	 */
484 	if (sort1 < sort2)
485 		return (-1);
486 	if (sort1 > sort2)
487 		return (1);
488 
489 	int cmp = AVL_CMP(m2->ms_weight, m1->ms_weight);
490 	if (likely(cmp))
491 		return (cmp);
492 
493 	IMPLY(AVL_CMP(m1->ms_start, m2->ms_start) == 0, m1 == m2);
494 
495 	return (AVL_CMP(m1->ms_start, m2->ms_start));
496 }
497 
498 uint64_t
499 metaslab_allocated_space(metaslab_t *msp)
500 {
501 	return (msp->ms_allocated_space);
502 }
503 
504 /*
505  * Verify that the space accounting on disk matches the in-core range_trees.
506  */
507 static void
508 metaslab_verify_space(metaslab_t *msp, uint64_t txg)
509 {
510 	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
511 	uint64_t allocating = 0;
512 	uint64_t sm_free_space, msp_free_space;
513 
514 	ASSERT(MUTEX_HELD(&msp->ms_lock));
515 	ASSERT(!msp->ms_condensing);
516 
517 	if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0)
518 		return;
519 
520 	/*
521 	 * We can only verify the metaslab space when we're called
522 	 * from syncing context with a loaded metaslab that has an
523 	 * allocated space map. Calling this in non-syncing context
524 	 * does not provide a consistent view of the metaslab since
525 	 * we're performing allocations in the future.
526 	 */
527 	if (txg != spa_syncing_txg(spa) || msp->ms_sm == NULL ||
528 	    !msp->ms_loaded)
529 		return;
530 
531 	/*
532 	 * Even though the smp_alloc field can get negative (e.g.
533 	 * see vdev_checkpoint_sm), that should never be the case
534 	 * when it come's to a metaslab's space map.
535 	 */
536 	ASSERT3S(space_map_allocated(msp->ms_sm), >=, 0);
537 
538 	sm_free_space = msp->ms_size - metaslab_allocated_space(msp);
539 
540 	/*
541 	 * Account for future allocations since we would have
542 	 * already deducted that space from the ms_allocatable.
543 	 */
544 	for (int t = 0; t < TXG_CONCURRENT_STATES; t++) {
545 		allocating +=
546 		    range_tree_space(msp->ms_allocating[(txg + t) & TXG_MASK]);
547 	}
548 
549 	ASSERT3U(msp->ms_deferspace, ==,
550 	    range_tree_space(msp->ms_defer[0]) +
551 	    range_tree_space(msp->ms_defer[1]));
552 
553 	msp_free_space = range_tree_space(msp->ms_allocatable) + allocating +
554 	    msp->ms_deferspace + range_tree_space(msp->ms_freed);
555 
556 	VERIFY3U(sm_free_space, ==, msp_free_space);
557 }
558 
559 /*
560  * ==========================================================================
561  * Metaslab groups
562  * ==========================================================================
563  */
564 /*
565  * Update the allocatable flag and the metaslab group's capacity.
566  * The allocatable flag is set to true if the capacity is below
567  * the zfs_mg_noalloc_threshold or has a fragmentation value that is
568  * greater than zfs_mg_fragmentation_threshold. If a metaslab group
569  * transitions from allocatable to non-allocatable or vice versa then the
570  * metaslab group's class is updated to reflect the transition.
571  */
572 static void
573 metaslab_group_alloc_update(metaslab_group_t *mg)
574 {
575 	vdev_t *vd = mg->mg_vd;
576 	metaslab_class_t *mc = mg->mg_class;
577 	vdev_stat_t *vs = &vd->vdev_stat;
578 	boolean_t was_allocatable;
579 	boolean_t was_initialized;
580 
581 	ASSERT(vd == vd->vdev_top);
582 	ASSERT3U(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_READER), ==,
583 	    SCL_ALLOC);
584 
585 	mutex_enter(&mg->mg_lock);
586 	was_allocatable = mg->mg_allocatable;
587 	was_initialized = mg->mg_initialized;
588 
589 	mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
590 	    (vs->vs_space + 1);
591 
592 	mutex_enter(&mc->mc_lock);
593 
594 	/*
595 	 * If the metaslab group was just added then it won't
596 	 * have any space until we finish syncing out this txg.
597 	 * At that point we will consider it initialized and available
598 	 * for allocations.  We also don't consider non-activated
599 	 * metaslab groups (e.g. vdevs that are in the middle of being removed)
600 	 * to be initialized, because they can't be used for allocation.
601 	 */
602 	mg->mg_initialized = metaslab_group_initialized(mg);
603 	if (!was_initialized && mg->mg_initialized) {
604 		mc->mc_groups++;
605 	} else if (was_initialized && !mg->mg_initialized) {
606 		ASSERT3U(mc->mc_groups, >, 0);
607 		mc->mc_groups--;
608 	}
609 	if (mg->mg_initialized)
610 		mg->mg_no_free_space = B_FALSE;
611 
612 	/*
613 	 * A metaslab group is considered allocatable if it has plenty
614 	 * of free space or is not heavily fragmented. We only take
615 	 * fragmentation into account if the metaslab group has a valid
616 	 * fragmentation metric (i.e. a value between 0 and 100).
617 	 */
618 	mg->mg_allocatable = (mg->mg_activation_count > 0 &&
619 	    mg->mg_free_capacity > zfs_mg_noalloc_threshold &&
620 	    (mg->mg_fragmentation == ZFS_FRAG_INVALID ||
621 	    mg->mg_fragmentation <= zfs_mg_fragmentation_threshold));
622 
623 	/*
624 	 * The mc_alloc_groups maintains a count of the number of
625 	 * groups in this metaslab class that are still above the
626 	 * zfs_mg_noalloc_threshold. This is used by the allocating
627 	 * threads to determine if they should avoid allocations to
628 	 * a given group. The allocator will avoid allocations to a group
629 	 * if that group has reached or is below the zfs_mg_noalloc_threshold
630 	 * and there are still other groups that are above the threshold.
631 	 * When a group transitions from allocatable to non-allocatable or
632 	 * vice versa we update the metaslab class to reflect that change.
633 	 * When the mc_alloc_groups value drops to 0 that means that all
634 	 * groups have reached the zfs_mg_noalloc_threshold making all groups
635 	 * eligible for allocations. This effectively means that all devices
636 	 * are balanced again.
637 	 */
638 	if (was_allocatable && !mg->mg_allocatable)
639 		mc->mc_alloc_groups--;
640 	else if (!was_allocatable && mg->mg_allocatable)
641 		mc->mc_alloc_groups++;
642 	mutex_exit(&mc->mc_lock);
643 
644 	mutex_exit(&mg->mg_lock);
645 }
646 
647 metaslab_group_t *
648 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd, int allocators)
649 {
650 	metaslab_group_t *mg;
651 
652 	mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
653 	mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
654 	mutex_init(&mg->mg_ms_disabled_lock, NULL, MUTEX_DEFAULT, NULL);
655 	cv_init(&mg->mg_ms_disabled_cv, NULL, CV_DEFAULT, NULL);
656 	mg->mg_primaries = kmem_zalloc(allocators * sizeof (metaslab_t *),
657 	    KM_SLEEP);
658 	mg->mg_secondaries = kmem_zalloc(allocators * sizeof (metaslab_t *),
659 	    KM_SLEEP);
660 	avl_create(&mg->mg_metaslab_tree, metaslab_compare,
661 	    sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
662 	mg->mg_vd = vd;
663 	mg->mg_class = mc;
664 	mg->mg_activation_count = 0;
665 	mg->mg_initialized = B_FALSE;
666 	mg->mg_no_free_space = B_TRUE;
667 	mg->mg_allocators = allocators;
668 
669 	mg->mg_alloc_queue_depth = kmem_zalloc(allocators *
670 	    sizeof (zfs_refcount_t), KM_SLEEP);
671 	mg->mg_cur_max_alloc_queue_depth = kmem_zalloc(allocators *
672 	    sizeof (uint64_t), KM_SLEEP);
673 	for (int i = 0; i < allocators; i++) {
674 		zfs_refcount_create_tracked(&mg->mg_alloc_queue_depth[i]);
675 		mg->mg_cur_max_alloc_queue_depth[i] = 0;
676 	}
677 
678 	mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct,
679 	    minclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT);
680 
681 	return (mg);
682 }
683 
684 void
685 metaslab_group_destroy(metaslab_group_t *mg)
686 {
687 	ASSERT(mg->mg_prev == NULL);
688 	ASSERT(mg->mg_next == NULL);
689 	/*
690 	 * We may have gone below zero with the activation count
691 	 * either because we never activated in the first place or
692 	 * because we're done, and possibly removing the vdev.
693 	 */
694 	ASSERT(mg->mg_activation_count <= 0);
695 
696 	taskq_destroy(mg->mg_taskq);
697 	avl_destroy(&mg->mg_metaslab_tree);
698 	kmem_free(mg->mg_primaries, mg->mg_allocators * sizeof (metaslab_t *));
699 	kmem_free(mg->mg_secondaries, mg->mg_allocators *
700 	    sizeof (metaslab_t *));
701 	mutex_destroy(&mg->mg_lock);
702 	mutex_destroy(&mg->mg_ms_disabled_lock);
703 	cv_destroy(&mg->mg_ms_disabled_cv);
704 
705 	for (int i = 0; i < mg->mg_allocators; i++) {
706 		zfs_refcount_destroy(&mg->mg_alloc_queue_depth[i]);
707 		mg->mg_cur_max_alloc_queue_depth[i] = 0;
708 	}
709 	kmem_free(mg->mg_alloc_queue_depth, mg->mg_allocators *
710 	    sizeof (zfs_refcount_t));
711 	kmem_free(mg->mg_cur_max_alloc_queue_depth, mg->mg_allocators *
712 	    sizeof (uint64_t));
713 
714 	kmem_free(mg, sizeof (metaslab_group_t));
715 }
716 
717 void
718 metaslab_group_activate(metaslab_group_t *mg)
719 {
720 	metaslab_class_t *mc = mg->mg_class;
721 	metaslab_group_t *mgprev, *mgnext;
722 
723 	ASSERT3U(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER), !=, 0);
724 
725 	ASSERT(mc->mc_rotor != mg);
726 	ASSERT(mg->mg_prev == NULL);
727 	ASSERT(mg->mg_next == NULL);
728 	ASSERT(mg->mg_activation_count <= 0);
729 
730 	if (++mg->mg_activation_count <= 0)
731 		return;
732 
733 	mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
734 	metaslab_group_alloc_update(mg);
735 
736 	if ((mgprev = mc->mc_rotor) == NULL) {
737 		mg->mg_prev = mg;
738 		mg->mg_next = mg;
739 	} else {
740 		mgnext = mgprev->mg_next;
741 		mg->mg_prev = mgprev;
742 		mg->mg_next = mgnext;
743 		mgprev->mg_next = mg;
744 		mgnext->mg_prev = mg;
745 	}
746 	mc->mc_rotor = mg;
747 }
748 
749 /*
750  * Passivate a metaslab group and remove it from the allocation rotor.
751  * Callers must hold both the SCL_ALLOC and SCL_ZIO lock prior to passivating
752  * a metaslab group. This function will momentarily drop spa_config_locks
753  * that are lower than the SCL_ALLOC lock (see comment below).
754  */
755 void
756 metaslab_group_passivate(metaslab_group_t *mg)
757 {
758 	metaslab_class_t *mc = mg->mg_class;
759 	spa_t *spa = mc->mc_spa;
760 	metaslab_group_t *mgprev, *mgnext;
761 	int locks = spa_config_held(spa, SCL_ALL, RW_WRITER);
762 
763 	ASSERT3U(spa_config_held(spa, SCL_ALLOC | SCL_ZIO, RW_WRITER), ==,
764 	    (SCL_ALLOC | SCL_ZIO));
765 
766 	if (--mg->mg_activation_count != 0) {
767 		ASSERT(mc->mc_rotor != mg);
768 		ASSERT(mg->mg_prev == NULL);
769 		ASSERT(mg->mg_next == NULL);
770 		ASSERT(mg->mg_activation_count < 0);
771 		return;
772 	}
773 
774 	/*
775 	 * The spa_config_lock is an array of rwlocks, ordered as
776 	 * follows (from highest to lowest):
777 	 *	SCL_CONFIG > SCL_STATE > SCL_L2ARC > SCL_ALLOC >
778 	 *	SCL_ZIO > SCL_FREE > SCL_VDEV
779 	 * (For more information about the spa_config_lock see spa_misc.c)
780 	 * The higher the lock, the broader its coverage. When we passivate
781 	 * a metaslab group, we must hold both the SCL_ALLOC and the SCL_ZIO
782 	 * config locks. However, the metaslab group's taskq might be trying
783 	 * to preload metaslabs so we must drop the SCL_ZIO lock and any
784 	 * lower locks to allow the I/O to complete. At a minimum,
785 	 * we continue to hold the SCL_ALLOC lock, which prevents any future
786 	 * allocations from taking place and any changes to the vdev tree.
787 	 */
788 	spa_config_exit(spa, locks & ~(SCL_ZIO - 1), spa);
789 	taskq_wait(mg->mg_taskq);
790 	spa_config_enter(spa, locks & ~(SCL_ZIO - 1), spa, RW_WRITER);
791 	metaslab_group_alloc_update(mg);
792 	for (int i = 0; i < mg->mg_allocators; i++) {
793 		metaslab_t *msp = mg->mg_primaries[i];
794 		if (msp != NULL) {
795 			mutex_enter(&msp->ms_lock);
796 			metaslab_passivate(msp,
797 			    metaslab_weight_from_range_tree(msp));
798 			mutex_exit(&msp->ms_lock);
799 		}
800 		msp = mg->mg_secondaries[i];
801 		if (msp != NULL) {
802 			mutex_enter(&msp->ms_lock);
803 			metaslab_passivate(msp,
804 			    metaslab_weight_from_range_tree(msp));
805 			mutex_exit(&msp->ms_lock);
806 		}
807 	}
808 
809 	mgprev = mg->mg_prev;
810 	mgnext = mg->mg_next;
811 
812 	if (mg == mgnext) {
813 		mc->mc_rotor = NULL;
814 	} else {
815 		mc->mc_rotor = mgnext;
816 		mgprev->mg_next = mgnext;
817 		mgnext->mg_prev = mgprev;
818 	}
819 
820 	mg->mg_prev = NULL;
821 	mg->mg_next = NULL;
822 }
823 
824 boolean_t
825 metaslab_group_initialized(metaslab_group_t *mg)
826 {
827 	vdev_t *vd = mg->mg_vd;
828 	vdev_stat_t *vs = &vd->vdev_stat;
829 
830 	return (vs->vs_space != 0 && mg->mg_activation_count > 0);
831 }
832 
833 uint64_t
834 metaslab_group_get_space(metaslab_group_t *mg)
835 {
836 	return ((1ULL << mg->mg_vd->vdev_ms_shift) * mg->mg_vd->vdev_ms_count);
837 }
838 
839 void
840 metaslab_group_histogram_verify(metaslab_group_t *mg)
841 {
842 	uint64_t *mg_hist;
843 	vdev_t *vd = mg->mg_vd;
844 	uint64_t ashift = vd->vdev_ashift;
845 	int i;
846 
847 	if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
848 		return;
849 
850 	mg_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
851 	    KM_SLEEP);
852 
853 	ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE, >=,
854 	    SPACE_MAP_HISTOGRAM_SIZE + ashift);
855 
856 	for (int m = 0; m < vd->vdev_ms_count; m++) {
857 		metaslab_t *msp = vd->vdev_ms[m];
858 		ASSERT(msp != NULL);
859 
860 		/* skip if not active or not a member */
861 		if (msp->ms_sm == NULL || msp->ms_group != mg)
862 			continue;
863 
864 		for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++)
865 			mg_hist[i + ashift] +=
866 			    msp->ms_sm->sm_phys->smp_histogram[i];
867 	}
868 
869 	for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i ++)
870 		VERIFY3U(mg_hist[i], ==, mg->mg_histogram[i]);
871 
872 	kmem_free(mg_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
873 }
874 
875 static void
876 metaslab_group_histogram_add(metaslab_group_t *mg, metaslab_t *msp)
877 {
878 	metaslab_class_t *mc = mg->mg_class;
879 	uint64_t ashift = mg->mg_vd->vdev_ashift;
880 
881 	ASSERT(MUTEX_HELD(&msp->ms_lock));
882 	if (msp->ms_sm == NULL)
883 		return;
884 
885 	mutex_enter(&mg->mg_lock);
886 	for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
887 		mg->mg_histogram[i + ashift] +=
888 		    msp->ms_sm->sm_phys->smp_histogram[i];
889 		mc->mc_histogram[i + ashift] +=
890 		    msp->ms_sm->sm_phys->smp_histogram[i];
891 	}
892 	mutex_exit(&mg->mg_lock);
893 }
894 
895 void
896 metaslab_group_histogram_remove(metaslab_group_t *mg, metaslab_t *msp)
897 {
898 	metaslab_class_t *mc = mg->mg_class;
899 	uint64_t ashift = mg->mg_vd->vdev_ashift;
900 
901 	ASSERT(MUTEX_HELD(&msp->ms_lock));
902 	if (msp->ms_sm == NULL)
903 		return;
904 
905 	mutex_enter(&mg->mg_lock);
906 	for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
907 		ASSERT3U(mg->mg_histogram[i + ashift], >=,
908 		    msp->ms_sm->sm_phys->smp_histogram[i]);
909 		ASSERT3U(mc->mc_histogram[i + ashift], >=,
910 		    msp->ms_sm->sm_phys->smp_histogram[i]);
911 
912 		mg->mg_histogram[i + ashift] -=
913 		    msp->ms_sm->sm_phys->smp_histogram[i];
914 		mc->mc_histogram[i + ashift] -=
915 		    msp->ms_sm->sm_phys->smp_histogram[i];
916 	}
917 	mutex_exit(&mg->mg_lock);
918 }
919 
920 static void
921 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
922 {
923 	ASSERT(msp->ms_group == NULL);
924 	mutex_enter(&mg->mg_lock);
925 	msp->ms_group = mg;
926 	msp->ms_weight = 0;
927 	avl_add(&mg->mg_metaslab_tree, msp);
928 	mutex_exit(&mg->mg_lock);
929 
930 	mutex_enter(&msp->ms_lock);
931 	metaslab_group_histogram_add(mg, msp);
932 	mutex_exit(&msp->ms_lock);
933 }
934 
935 static void
936 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
937 {
938 	mutex_enter(&msp->ms_lock);
939 	metaslab_group_histogram_remove(mg, msp);
940 	mutex_exit(&msp->ms_lock);
941 
942 	mutex_enter(&mg->mg_lock);
943 	ASSERT(msp->ms_group == mg);
944 	avl_remove(&mg->mg_metaslab_tree, msp);
945 	msp->ms_group = NULL;
946 	mutex_exit(&mg->mg_lock);
947 }
948 
949 static void
950 metaslab_group_sort_impl(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
951 {
952 	ASSERT(MUTEX_HELD(&mg->mg_lock));
953 	ASSERT(msp->ms_group == mg);
954 	avl_remove(&mg->mg_metaslab_tree, msp);
955 	msp->ms_weight = weight;
956 	avl_add(&mg->mg_metaslab_tree, msp);
957 
958 }
959 
960 static void
961 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
962 {
963 	/*
964 	 * Although in principle the weight can be any value, in
965 	 * practice we do not use values in the range [1, 511].
966 	 */
967 	ASSERT(weight >= SPA_MINBLOCKSIZE || weight == 0);
968 	ASSERT(MUTEX_HELD(&msp->ms_lock));
969 
970 	mutex_enter(&mg->mg_lock);
971 	metaslab_group_sort_impl(mg, msp, weight);
972 	mutex_exit(&mg->mg_lock);
973 }
974 
975 /*
976  * Calculate the fragmentation for a given metaslab group. We can use
977  * a simple average here since all metaslabs within the group must have
978  * the same size. The return value will be a value between 0 and 100
979  * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
980  * group have a fragmentation metric.
981  */
982 uint64_t
983 metaslab_group_fragmentation(metaslab_group_t *mg)
984 {
985 	vdev_t *vd = mg->mg_vd;
986 	uint64_t fragmentation = 0;
987 	uint64_t valid_ms = 0;
988 
989 	for (int m = 0; m < vd->vdev_ms_count; m++) {
990 		metaslab_t *msp = vd->vdev_ms[m];
991 
992 		if (msp->ms_fragmentation == ZFS_FRAG_INVALID)
993 			continue;
994 		if (msp->ms_group != mg)
995 			continue;
996 
997 		valid_ms++;
998 		fragmentation += msp->ms_fragmentation;
999 	}
1000 
1001 	if (valid_ms <= mg->mg_vd->vdev_ms_count / 2)
1002 		return (ZFS_FRAG_INVALID);
1003 
1004 	fragmentation /= valid_ms;
1005 	ASSERT3U(fragmentation, <=, 100);
1006 	return (fragmentation);
1007 }
1008 
1009 /*
1010  * Determine if a given metaslab group should skip allocations. A metaslab
1011  * group should avoid allocations if its free capacity is less than the
1012  * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
1013  * zfs_mg_fragmentation_threshold and there is at least one metaslab group
1014  * that can still handle allocations. If the allocation throttle is enabled
1015  * then we skip allocations to devices that have reached their maximum
1016  * allocation queue depth unless the selected metaslab group is the only
1017  * eligible group remaining.
1018  */
1019 static boolean_t
1020 metaslab_group_allocatable(metaslab_group_t *mg, metaslab_group_t *rotor,
1021     uint64_t psize, int allocator, int d)
1022 {
1023 	spa_t *spa = mg->mg_vd->vdev_spa;
1024 	metaslab_class_t *mc = mg->mg_class;
1025 
1026 	/*
1027 	 * We can only consider skipping this metaslab group if it's
1028 	 * in the normal metaslab class and there are other metaslab
1029 	 * groups to select from. Otherwise, we always consider it eligible
1030 	 * for allocations.
1031 	 */
1032 	if ((mc != spa_normal_class(spa) &&
1033 	    mc != spa_special_class(spa) &&
1034 	    mc != spa_dedup_class(spa)) ||
1035 	    mc->mc_groups <= 1)
1036 		return (B_TRUE);
1037 
1038 	/*
1039 	 * If the metaslab group's mg_allocatable flag is set (see comments
1040 	 * in metaslab_group_alloc_update() for more information) and
1041 	 * the allocation throttle is disabled then allow allocations to this
1042 	 * device. However, if the allocation throttle is enabled then
1043 	 * check if we have reached our allocation limit (mg_alloc_queue_depth)
1044 	 * to determine if we should allow allocations to this metaslab group.
1045 	 * If all metaslab groups are no longer considered allocatable
1046 	 * (mc_alloc_groups == 0) or we're trying to allocate the smallest
1047 	 * gang block size then we allow allocations on this metaslab group
1048 	 * regardless of the mg_allocatable or throttle settings.
1049 	 */
1050 	if (mg->mg_allocatable) {
1051 		metaslab_group_t *mgp;
1052 		int64_t qdepth;
1053 		uint64_t qmax = mg->mg_cur_max_alloc_queue_depth[allocator];
1054 
1055 		if (!mc->mc_alloc_throttle_enabled)
1056 			return (B_TRUE);
1057 
1058 		/*
1059 		 * If this metaslab group does not have any free space, then
1060 		 * there is no point in looking further.
1061 		 */
1062 		if (mg->mg_no_free_space)
1063 			return (B_FALSE);
1064 
1065 		/*
1066 		 * Relax allocation throttling for ditto blocks.  Due to
1067 		 * random imbalances in allocation it tends to push copies
1068 		 * to one vdev, that looks a bit better at the moment.
1069 		 */
1070 		qmax = qmax * (4 + d) / 4;
1071 
1072 		qdepth = zfs_refcount_count(
1073 		    &mg->mg_alloc_queue_depth[allocator]);
1074 
1075 		/*
1076 		 * If this metaslab group is below its qmax or it's
1077 		 * the only allocatable metasable group, then attempt
1078 		 * to allocate from it.
1079 		 */
1080 		if (qdepth < qmax || mc->mc_alloc_groups == 1)
1081 			return (B_TRUE);
1082 		ASSERT3U(mc->mc_alloc_groups, >, 1);
1083 
1084 		/*
1085 		 * Since this metaslab group is at or over its qmax, we
1086 		 * need to determine if there are metaslab groups after this
1087 		 * one that might be able to handle this allocation. This is
1088 		 * racy since we can't hold the locks for all metaslab
1089 		 * groups at the same time when we make this check.
1090 		 */
1091 		for (mgp = mg->mg_next; mgp != rotor; mgp = mgp->mg_next) {
1092 			qmax = mgp->mg_cur_max_alloc_queue_depth[allocator];
1093 			qmax = qmax * (4 + d) / 4;
1094 			qdepth = zfs_refcount_count(
1095 			    &mgp->mg_alloc_queue_depth[allocator]);
1096 
1097 			/*
1098 			 * If there is another metaslab group that
1099 			 * might be able to handle the allocation, then
1100 			 * we return false so that we skip this group.
1101 			 */
1102 			if (qdepth < qmax && !mgp->mg_no_free_space)
1103 				return (B_FALSE);
1104 		}
1105 
1106 		/*
1107 		 * We didn't find another group to handle the allocation
1108 		 * so we can't skip this metaslab group even though
1109 		 * we are at or over our qmax.
1110 		 */
1111 		return (B_TRUE);
1112 
1113 	} else if (mc->mc_alloc_groups == 0 || psize == SPA_MINBLOCKSIZE) {
1114 		return (B_TRUE);
1115 	}
1116 	return (B_FALSE);
1117 }
1118 
1119 /*
1120  * ==========================================================================
1121  * Range tree callbacks
1122  * ==========================================================================
1123  */
1124 
1125 /*
1126  * Comparison function for the private size-ordered tree. Tree is sorted
1127  * by size, larger sizes at the end of the tree.
1128  */
1129 static int
1130 metaslab_rangesize_compare(const void *x1, const void *x2)
1131 {
1132 	const range_seg_t *r1 = x1;
1133 	const range_seg_t *r2 = x2;
1134 	uint64_t rs_size1 = r1->rs_end - r1->rs_start;
1135 	uint64_t rs_size2 = r2->rs_end - r2->rs_start;
1136 
1137 	int cmp = AVL_CMP(rs_size1, rs_size2);
1138 	if (likely(cmp))
1139 		return (cmp);
1140 
1141 	return (AVL_CMP(r1->rs_start, r2->rs_start));
1142 }
1143 
1144 /*
1145  * ==========================================================================
1146  * Common allocator routines
1147  * ==========================================================================
1148  */
1149 
1150 /*
1151  * Return the maximum contiguous segment within the metaslab.
1152  */
1153 uint64_t
1154 metaslab_block_maxsize(metaslab_t *msp)
1155 {
1156 	avl_tree_t *t = &msp->ms_allocatable_by_size;
1157 	range_seg_t *rs;
1158 
1159 	if (t == NULL || (rs = avl_last(t)) == NULL)
1160 		return (0ULL);
1161 
1162 	return (rs->rs_end - rs->rs_start);
1163 }
1164 
1165 static range_seg_t *
1166 metaslab_block_find(avl_tree_t *t, uint64_t start, uint64_t size)
1167 {
1168 	range_seg_t *rs, rsearch;
1169 	avl_index_t where;
1170 
1171 	rsearch.rs_start = start;
1172 	rsearch.rs_end = start + size;
1173 
1174 	rs = avl_find(t, &rsearch, &where);
1175 	if (rs == NULL) {
1176 		rs = avl_nearest(t, where, AVL_AFTER);
1177 	}
1178 
1179 	return (rs);
1180 }
1181 
1182 /*
1183  * This is a helper function that can be used by the allocator to find
1184  * a suitable block to allocate. This will search the specified AVL
1185  * tree looking for a block that matches the specified criteria.
1186  */
1187 static uint64_t
1188 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
1189     uint64_t align)
1190 {
1191 	range_seg_t *rs = metaslab_block_find(t, *cursor, size);
1192 
1193 	while (rs != NULL) {
1194 		uint64_t offset = P2ROUNDUP(rs->rs_start, align);
1195 
1196 		if (offset + size <= rs->rs_end) {
1197 			*cursor = offset + size;
1198 			return (offset);
1199 		}
1200 		rs = AVL_NEXT(t, rs);
1201 	}
1202 
1203 	/*
1204 	 * If we know we've searched the whole map (*cursor == 0), give up.
1205 	 * Otherwise, reset the cursor to the beginning and try again.
1206 	 */
1207 	if (*cursor == 0)
1208 		return (-1ULL);
1209 
1210 	*cursor = 0;
1211 	return (metaslab_block_picker(t, cursor, size, align));
1212 }
1213 
1214 /*
1215  * ==========================================================================
1216  * The first-fit block allocator
1217  * ==========================================================================
1218  */
1219 static uint64_t
1220 metaslab_ff_alloc(metaslab_t *msp, uint64_t size)
1221 {
1222 	/*
1223 	 * Find the largest power of 2 block size that evenly divides the
1224 	 * requested size. This is used to try to allocate blocks with similar
1225 	 * alignment from the same area of the metaslab (i.e. same cursor
1226 	 * bucket) but it does not guarantee that other allocations sizes
1227 	 * may exist in the same region.
1228 	 */
1229 	uint64_t align = size & -size;
1230 	uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1231 	avl_tree_t *t = &msp->ms_allocatable->rt_root;
1232 
1233 	return (metaslab_block_picker(t, cursor, size, align));
1234 }
1235 
1236 static metaslab_ops_t metaslab_ff_ops = {
1237 	metaslab_ff_alloc
1238 };
1239 
1240 /*
1241  * ==========================================================================
1242  * Dynamic block allocator -
1243  * Uses the first fit allocation scheme until space get low and then
1244  * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1245  * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1246  * ==========================================================================
1247  */
1248 static uint64_t
1249 metaslab_df_alloc(metaslab_t *msp, uint64_t size)
1250 {
1251 	/*
1252 	 * Find the largest power of 2 block size that evenly divides the
1253 	 * requested size. This is used to try to allocate blocks with similar
1254 	 * alignment from the same area of the metaslab (i.e. same cursor
1255 	 * bucket) but it does not guarantee that other allocations sizes
1256 	 * may exist in the same region.
1257 	 */
1258 	uint64_t align = size & -size;
1259 	uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1260 	range_tree_t *rt = msp->ms_allocatable;
1261 	avl_tree_t *t = &rt->rt_root;
1262 	uint64_t max_size = metaslab_block_maxsize(msp);
1263 	int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
1264 
1265 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1266 	ASSERT3U(avl_numnodes(t), ==,
1267 	    avl_numnodes(&msp->ms_allocatable_by_size));
1268 
1269 	if (max_size < size)
1270 		return (-1ULL);
1271 
1272 	/*
1273 	 * If we're running low on space switch to using the size
1274 	 * sorted AVL tree (best-fit).
1275 	 */
1276 	if (max_size < metaslab_df_alloc_threshold ||
1277 	    free_pct < metaslab_df_free_pct) {
1278 		t = &msp->ms_allocatable_by_size;
1279 		*cursor = 0;
1280 	}
1281 
1282 	return (metaslab_block_picker(t, cursor, size, 1ULL));
1283 }
1284 
1285 static metaslab_ops_t metaslab_df_ops = {
1286 	metaslab_df_alloc
1287 };
1288 
1289 /*
1290  * ==========================================================================
1291  * Cursor fit block allocator -
1292  * Select the largest region in the metaslab, set the cursor to the beginning
1293  * of the range and the cursor_end to the end of the range. As allocations
1294  * are made advance the cursor. Continue allocating from the cursor until
1295  * the range is exhausted and then find a new range.
1296  * ==========================================================================
1297  */
1298 static uint64_t
1299 metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
1300 {
1301 	range_tree_t *rt = msp->ms_allocatable;
1302 	avl_tree_t *t = &msp->ms_allocatable_by_size;
1303 	uint64_t *cursor = &msp->ms_lbas[0];
1304 	uint64_t *cursor_end = &msp->ms_lbas[1];
1305 	uint64_t offset = 0;
1306 
1307 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1308 	ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&rt->rt_root));
1309 
1310 	ASSERT3U(*cursor_end, >=, *cursor);
1311 
1312 	if ((*cursor + size) > *cursor_end) {
1313 		range_seg_t *rs;
1314 
1315 		rs = avl_last(&msp->ms_allocatable_by_size);
1316 		if (rs == NULL || (rs->rs_end - rs->rs_start) < size)
1317 			return (-1ULL);
1318 
1319 		*cursor = rs->rs_start;
1320 		*cursor_end = rs->rs_end;
1321 	}
1322 
1323 	offset = *cursor;
1324 	*cursor += size;
1325 
1326 	return (offset);
1327 }
1328 
1329 static metaslab_ops_t metaslab_cf_ops = {
1330 	metaslab_cf_alloc
1331 };
1332 
1333 /*
1334  * ==========================================================================
1335  * New dynamic fit allocator -
1336  * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1337  * contiguous blocks. If no region is found then just use the largest segment
1338  * that remains.
1339  * ==========================================================================
1340  */
1341 
1342 /*
1343  * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1344  * to request from the allocator.
1345  */
1346 uint64_t metaslab_ndf_clump_shift = 4;
1347 
1348 static uint64_t
1349 metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
1350 {
1351 	avl_tree_t *t = &msp->ms_allocatable->rt_root;
1352 	avl_index_t where;
1353 	range_seg_t *rs, rsearch;
1354 	uint64_t hbit = highbit64(size);
1355 	uint64_t *cursor = &msp->ms_lbas[hbit - 1];
1356 	uint64_t max_size = metaslab_block_maxsize(msp);
1357 
1358 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1359 	ASSERT3U(avl_numnodes(t), ==,
1360 	    avl_numnodes(&msp->ms_allocatable_by_size));
1361 
1362 	if (max_size < size)
1363 		return (-1ULL);
1364 
1365 	rsearch.rs_start = *cursor;
1366 	rsearch.rs_end = *cursor + size;
1367 
1368 	rs = avl_find(t, &rsearch, &where);
1369 	if (rs == NULL || (rs->rs_end - rs->rs_start) < size) {
1370 		t = &msp->ms_allocatable_by_size;
1371 
1372 		rsearch.rs_start = 0;
1373 		rsearch.rs_end = MIN(max_size,
1374 		    1ULL << (hbit + metaslab_ndf_clump_shift));
1375 		rs = avl_find(t, &rsearch, &where);
1376 		if (rs == NULL)
1377 			rs = avl_nearest(t, where, AVL_AFTER);
1378 		ASSERT(rs != NULL);
1379 	}
1380 
1381 	if ((rs->rs_end - rs->rs_start) >= size) {
1382 		*cursor = rs->rs_start + size;
1383 		return (rs->rs_start);
1384 	}
1385 	return (-1ULL);
1386 }
1387 
1388 static metaslab_ops_t metaslab_ndf_ops = {
1389 	metaslab_ndf_alloc
1390 };
1391 
1392 metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
1393 
1394 /*
1395  * ==========================================================================
1396  * Metaslabs
1397  * ==========================================================================
1398  */
1399 
1400 static void
1401 metaslab_aux_histograms_clear(metaslab_t *msp)
1402 {
1403 	/*
1404 	 * Auxiliary histograms are only cleared when resetting them,
1405 	 * which can only happen while the metaslab is loaded.
1406 	 */
1407 	ASSERT(msp->ms_loaded);
1408 
1409 	bzero(msp->ms_synchist, sizeof (msp->ms_synchist));
1410 	for (int t = 0; t < TXG_DEFER_SIZE; t++)
1411 		bzero(msp->ms_deferhist[t], sizeof (msp->ms_deferhist[t]));
1412 }
1413 
1414 static void
1415 metaslab_aux_histogram_add(uint64_t *histogram, uint64_t shift,
1416     range_tree_t *rt)
1417 {
1418 	/*
1419 	 * This is modeled after space_map_histogram_add(), so refer to that
1420 	 * function for implementation details. We want this to work like
1421 	 * the space map histogram, and not the range tree histogram, as we
1422 	 * are essentially constructing a delta that will be later subtracted
1423 	 * from the space map histogram.
1424 	 */
1425 	int idx = 0;
1426 	for (int i = shift; i < RANGE_TREE_HISTOGRAM_SIZE; i++) {
1427 		ASSERT3U(i, >=, idx + shift);
1428 		histogram[idx] += rt->rt_histogram[i] << (i - idx - shift);
1429 
1430 		if (idx < SPACE_MAP_HISTOGRAM_SIZE - 1) {
1431 			ASSERT3U(idx + shift, ==, i);
1432 			idx++;
1433 			ASSERT3U(idx, <, SPACE_MAP_HISTOGRAM_SIZE);
1434 		}
1435 	}
1436 }
1437 
1438 /*
1439  * Called at every sync pass that the metaslab gets synced.
1440  *
1441  * The reason is that we want our auxiliary histograms to be updated
1442  * wherever the metaslab's space map histogram is updated. This way
1443  * we stay consistent on which parts of the metaslab space map's
1444  * histogram are currently not available for allocations (e.g because
1445  * they are in the defer, freed, and freeing trees).
1446  */
1447 static void
1448 metaslab_aux_histograms_update(metaslab_t *msp)
1449 {
1450 	space_map_t *sm = msp->ms_sm;
1451 	ASSERT(sm != NULL);
1452 
1453 	/*
1454 	 * This is similar to the metaslab's space map histogram updates
1455 	 * that take place in metaslab_sync(). The only difference is that
1456 	 * we only care about segments that haven't made it into the
1457 	 * ms_allocatable tree yet.
1458 	 */
1459 	if (msp->ms_loaded) {
1460 		metaslab_aux_histograms_clear(msp);
1461 
1462 		metaslab_aux_histogram_add(msp->ms_synchist,
1463 		    sm->sm_shift, msp->ms_freed);
1464 
1465 		for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1466 			metaslab_aux_histogram_add(msp->ms_deferhist[t],
1467 			    sm->sm_shift, msp->ms_defer[t]);
1468 		}
1469 	}
1470 
1471 	metaslab_aux_histogram_add(msp->ms_synchist,
1472 	    sm->sm_shift, msp->ms_freeing);
1473 }
1474 
1475 /*
1476  * Called every time we are done syncing (writing to) the metaslab,
1477  * i.e. at the end of each sync pass.
1478  * [see the comment in metaslab_impl.h for ms_synchist, ms_deferhist]
1479  */
1480 static void
1481 metaslab_aux_histograms_update_done(metaslab_t *msp, boolean_t defer_allowed)
1482 {
1483 	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1484 	space_map_t *sm = msp->ms_sm;
1485 
1486 	if (sm == NULL) {
1487 		/*
1488 		 * We came here from metaslab_init() when creating/opening a
1489 		 * pool, looking at a metaslab that hasn't had any allocations
1490 		 * yet.
1491 		 */
1492 		return;
1493 	}
1494 
1495 	/*
1496 	 * This is similar to the actions that we take for the ms_freed
1497 	 * and ms_defer trees in metaslab_sync_done().
1498 	 */
1499 	uint64_t hist_index = spa_syncing_txg(spa) % TXG_DEFER_SIZE;
1500 	if (defer_allowed) {
1501 		bcopy(msp->ms_synchist, msp->ms_deferhist[hist_index],
1502 		    sizeof (msp->ms_synchist));
1503 	} else {
1504 		bzero(msp->ms_deferhist[hist_index],
1505 		    sizeof (msp->ms_deferhist[hist_index]));
1506 	}
1507 	bzero(msp->ms_synchist, sizeof (msp->ms_synchist));
1508 }
1509 
1510 /*
1511  * Ensure that the metaslab's weight and fragmentation are consistent
1512  * with the contents of the histogram (either the range tree's histogram
1513  * or the space map's depending whether the metaslab is loaded).
1514  */
1515 static void
1516 metaslab_verify_weight_and_frag(metaslab_t *msp)
1517 {
1518 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1519 
1520 	if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0)
1521 		return;
1522 
1523 	/* see comment in metaslab_verify_unflushed_changes() */
1524 	if (msp->ms_group == NULL)
1525 		return;
1526 
1527 	/*
1528 	 * Devices being removed always return a weight of 0 and leave
1529 	 * fragmentation and ms_max_size as is - there is nothing for
1530 	 * us to verify here.
1531 	 */
1532 	vdev_t *vd = msp->ms_group->mg_vd;
1533 	if (vd->vdev_removing)
1534 		return;
1535 
1536 	/*
1537 	 * If the metaslab is dirty it probably means that we've done
1538 	 * some allocations or frees that have changed our histograms
1539 	 * and thus the weight.
1540 	 */
1541 	for (int t = 0; t < TXG_SIZE; t++) {
1542 		if (txg_list_member(&vd->vdev_ms_list, msp, t))
1543 			return;
1544 	}
1545 
1546 	/*
1547 	 * This verification checks that our in-memory state is consistent
1548 	 * with what's on disk. If the pool is read-only then there aren't
1549 	 * any changes and we just have the initially-loaded state.
1550 	 */
1551 	if (!spa_writeable(msp->ms_group->mg_vd->vdev_spa))
1552 		return;
1553 
1554 	/* some extra verification for in-core tree if you can */
1555 	if (msp->ms_loaded) {
1556 		range_tree_stat_verify(msp->ms_allocatable);
1557 		VERIFY(space_map_histogram_verify(msp->ms_sm,
1558 		    msp->ms_allocatable));
1559 	}
1560 
1561 	uint64_t weight = msp->ms_weight;
1562 	uint64_t was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
1563 	boolean_t space_based = WEIGHT_IS_SPACEBASED(msp->ms_weight);
1564 	uint64_t frag = msp->ms_fragmentation;
1565 	uint64_t max_segsize = msp->ms_max_size;
1566 
1567 	msp->ms_weight = 0;
1568 	msp->ms_fragmentation = 0;
1569 	msp->ms_max_size = 0;
1570 
1571 	/*
1572 	 * This function is used for verification purposes. Regardless of
1573 	 * whether metaslab_weight() thinks this metaslab should be active or
1574 	 * not, we want to ensure that the actual weight (and therefore the
1575 	 * value of ms_weight) would be the same if it was to be recalculated
1576 	 * at this point.
1577 	 */
1578 	msp->ms_weight = metaslab_weight(msp) | was_active;
1579 
1580 	VERIFY3U(max_segsize, ==, msp->ms_max_size);
1581 
1582 	/*
1583 	 * If the weight type changed then there is no point in doing
1584 	 * verification. Revert fields to their original values.
1585 	 */
1586 	if ((space_based && !WEIGHT_IS_SPACEBASED(msp->ms_weight)) ||
1587 	    (!space_based && WEIGHT_IS_SPACEBASED(msp->ms_weight))) {
1588 		msp->ms_fragmentation = frag;
1589 		msp->ms_weight = weight;
1590 		return;
1591 	}
1592 
1593 	VERIFY3U(msp->ms_fragmentation, ==, frag);
1594 	VERIFY3U(msp->ms_weight, ==, weight);
1595 }
1596 
1597 /*
1598  * Wait for any in-progress metaslab loads to complete.
1599  */
1600 static void
1601 metaslab_load_wait(metaslab_t *msp)
1602 {
1603 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1604 
1605 	while (msp->ms_loading) {
1606 		ASSERT(!msp->ms_loaded);
1607 		cv_wait(&msp->ms_load_cv, &msp->ms_lock);
1608 	}
1609 }
1610 
1611 static int
1612 metaslab_load_impl(metaslab_t *msp)
1613 {
1614 	int error = 0;
1615 
1616 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1617 	ASSERT(msp->ms_loading);
1618 	ASSERT(!msp->ms_condensing);
1619 
1620 	/*
1621 	 * We temporarily drop the lock to unblock other operations while we
1622 	 * are reading the space map. Therefore, metaslab_sync() and
1623 	 * metaslab_sync_done() can run at the same time as we do.
1624 	 *
1625 	 * metaslab_sync() can append to the space map while we are loading.
1626 	 * Therefore we load only entries that existed when we started the
1627 	 * load. Additionally, metaslab_sync_done() has to wait for the load
1628 	 * to complete because there are potential races like metaslab_load()
1629 	 * loading parts of the space map that are currently being appended
1630 	 * by metaslab_sync(). If we didn't, the ms_allocatable would have
1631 	 * entries that metaslab_sync_done() would try to re-add later.
1632 	 *
1633 	 * That's why before dropping the lock we remember the synced length
1634 	 * of the metaslab and read up to that point of the space map,
1635 	 * ignoring entries appended by metaslab_sync() that happen after we
1636 	 * drop the lock.
1637 	 */
1638 	uint64_t length = msp->ms_synced_length;
1639 	mutex_exit(&msp->ms_lock);
1640 
1641 	if (msp->ms_sm != NULL) {
1642 		error = space_map_load_length(msp->ms_sm, msp->ms_allocatable,
1643 		    SM_FREE, length);
1644 	} else {
1645 		/*
1646 		 * The space map has not been allocated yet, so treat
1647 		 * all the space in the metaslab as free and add it to the
1648 		 * ms_allocatable tree.
1649 		 */
1650 		range_tree_add(msp->ms_allocatable,
1651 		    msp->ms_start, msp->ms_size);
1652 	}
1653 
1654 	/*
1655 	 * We need to grab the ms_sync_lock to prevent metaslab_sync() from
1656 	 * changing the ms_sm and the metaslab's range trees while we are
1657 	 * about to use them and populate the ms_allocatable. The ms_lock
1658 	 * is insufficient for this because metaslab_sync() doesn't hold
1659 	 * the ms_lock while writing the ms_checkpointing tree to disk.
1660 	 */
1661 	mutex_enter(&msp->ms_sync_lock);
1662 	mutex_enter(&msp->ms_lock);
1663 	ASSERT(!msp->ms_condensing);
1664 
1665 	if (error != 0) {
1666 		mutex_exit(&msp->ms_sync_lock);
1667 		return (error);
1668 	}
1669 
1670 	ASSERT3P(msp->ms_group, !=, NULL);
1671 	msp->ms_loaded = B_TRUE;
1672 
1673 	/*
1674 	 * The ms_allocatable contains the segments that exist in the
1675 	 * ms_defer trees [see ms_synced_length]. Thus we need to remove
1676 	 * them from ms_allocatable as they will be added again in
1677 	 * metaslab_sync_done().
1678 	 */
1679 	for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1680 		range_tree_walk(msp->ms_defer[t],
1681 		    range_tree_remove, msp->ms_allocatable);
1682 	}
1683 
1684 	/*
1685 	 * Call metaslab_recalculate_weight_and_sort() now that the
1686 	 * metaslab is loaded so we get the metaslab's real weight.
1687 	 *
1688 	 * Unless this metaslab was created with older software and
1689 	 * has not yet been converted to use segment-based weight, we
1690 	 * expect the new weight to be better or equal to the weight
1691 	 * that the metaslab had while it was not loaded. This is
1692 	 * because the old weight does not take into account the
1693 	 * consolidation of adjacent segments between TXGs. [see
1694 	 * comment for ms_synchist and ms_deferhist[] for more info]
1695 	 */
1696 	uint64_t weight = msp->ms_weight;
1697 	metaslab_recalculate_weight_and_sort(msp);
1698 	if (!WEIGHT_IS_SPACEBASED(weight))
1699 		ASSERT3U(weight, <=, msp->ms_weight);
1700 	msp->ms_max_size = metaslab_block_maxsize(msp);
1701 
1702 	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1703 	metaslab_verify_space(msp, spa_syncing_txg(spa));
1704 	mutex_exit(&msp->ms_sync_lock);
1705 
1706 	return (0);
1707 }
1708 
1709 int
1710 metaslab_load(metaslab_t *msp)
1711 {
1712 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1713 
1714 	/*
1715 	 * There may be another thread loading the same metaslab, if that's
1716 	 * the case just wait until the other thread is done and return.
1717 	 */
1718 	metaslab_load_wait(msp);
1719 	if (msp->ms_loaded)
1720 		return (0);
1721 	VERIFY(!msp->ms_loading);
1722 	ASSERT(!msp->ms_condensing);
1723 
1724 	msp->ms_loading = B_TRUE;
1725 	int error = metaslab_load_impl(msp);
1726 	msp->ms_loading = B_FALSE;
1727 	cv_broadcast(&msp->ms_load_cv);
1728 
1729 	return (error);
1730 }
1731 
1732 void
1733 metaslab_unload(metaslab_t *msp)
1734 {
1735 	ASSERT(MUTEX_HELD(&msp->ms_lock));
1736 
1737 	metaslab_verify_weight_and_frag(msp);
1738 
1739 	range_tree_vacate(msp->ms_allocatable, NULL, NULL);
1740 	msp->ms_loaded = B_FALSE;
1741 
1742 	msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
1743 	msp->ms_max_size = 0;
1744 
1745 	/*
1746 	 * We explicitly recalculate the metaslab's weight based on its space
1747 	 * map (as it is now not loaded). We want unload metaslabs to always
1748 	 * have their weights calculated from the space map histograms, while
1749 	 * loaded ones have it calculated from their in-core range tree
1750 	 * [see metaslab_load()]. This way, the weight reflects the information
1751 	 * available in-core, whether it is loaded or not
1752 	 *
1753 	 * If ms_group == NULL means that we came here from metaslab_fini(),
1754 	 * at which point it doesn't make sense for us to do the recalculation
1755 	 * and the sorting.
1756 	 */
1757 	if (msp->ms_group != NULL)
1758 		metaslab_recalculate_weight_and_sort(msp);
1759 }
1760 
1761 static void
1762 metaslab_space_update(vdev_t *vd, metaslab_class_t *mc, int64_t alloc_delta,
1763     int64_t defer_delta, int64_t space_delta)
1764 {
1765 	vdev_space_update(vd, alloc_delta, defer_delta, space_delta);
1766 
1767 	ASSERT3P(vd->vdev_spa->spa_root_vdev, ==, vd->vdev_parent);
1768 	ASSERT(vd->vdev_ms_count != 0);
1769 
1770 	metaslab_class_space_update(mc, alloc_delta, defer_delta, space_delta,
1771 	    vdev_deflated_space(vd, space_delta));
1772 }
1773 
1774 int
1775 metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, uint64_t txg,
1776     metaslab_t **msp)
1777 {
1778 	vdev_t *vd = mg->mg_vd;
1779 	spa_t *spa = vd->vdev_spa;
1780 	objset_t *mos = spa->spa_meta_objset;
1781 	metaslab_t *ms;
1782 	int error;
1783 
1784 	ms = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
1785 	mutex_init(&ms->ms_lock, NULL, MUTEX_DEFAULT, NULL);
1786 	mutex_init(&ms->ms_sync_lock, NULL, MUTEX_DEFAULT, NULL);
1787 	cv_init(&ms->ms_load_cv, NULL, CV_DEFAULT, NULL);
1788 
1789 	ms->ms_id = id;
1790 	ms->ms_start = id << vd->vdev_ms_shift;
1791 	ms->ms_size = 1ULL << vd->vdev_ms_shift;
1792 	ms->ms_allocator = -1;
1793 	ms->ms_new = B_TRUE;
1794 
1795 	/*
1796 	 * We only open space map objects that already exist. All others
1797 	 * will be opened when we finally allocate an object for it.
1798 	 *
1799 	 * Note:
1800 	 * When called from vdev_expand(), we can't call into the DMU as
1801 	 * we are holding the spa_config_lock as a writer and we would
1802 	 * deadlock [see relevant comment in vdev_metaslab_init()]. in
1803 	 * that case, the object parameter is zero though, so we won't
1804 	 * call into the DMU.
1805 	 */
1806 	if (object != 0) {
1807 		error = space_map_open(&ms->ms_sm, mos, object, ms->ms_start,
1808 		    ms->ms_size, vd->vdev_ashift);
1809 
1810 		if (error != 0) {
1811 			kmem_free(ms, sizeof (metaslab_t));
1812 			return (error);
1813 		}
1814 
1815 		ASSERT(ms->ms_sm != NULL);
1816 		ASSERT3S(space_map_allocated(ms->ms_sm), >=, 0);
1817 		ms->ms_allocated_space = space_map_allocated(ms->ms_sm);
1818 	}
1819 
1820 	/*
1821 	 * We create the ms_allocatable here, but we don't create the
1822 	 * other range trees until metaslab_sync_done().  This serves
1823 	 * two purposes: it allows metaslab_sync_done() to detect the
1824 	 * addition of new space; and for debugging, it ensures that
1825 	 * we'd data fault on any attempt to use this metaslab before
1826 	 * it's ready.
1827 	 */
1828 	ms->ms_allocatable = range_tree_create_impl(&rt_avl_ops,
1829 	    &ms->ms_allocatable_by_size, metaslab_rangesize_compare, 0);
1830 
1831 	ms->ms_trim = range_tree_create(NULL, NULL);
1832 
1833 	metaslab_group_add(mg, ms);
1834 	metaslab_set_fragmentation(ms);
1835 
1836 	/*
1837 	 * If we're opening an existing pool (txg == 0) or creating
1838 	 * a new one (txg == TXG_INITIAL), all space is available now.
1839 	 * If we're adding space to an existing pool, the new space
1840 	 * does not become available until after this txg has synced.
1841 	 * The metaslab's weight will also be initialized when we sync
1842 	 * out this txg. This ensures that we don't attempt to allocate
1843 	 * from it before we have initialized it completely.
1844 	 */
1845 	if (txg <= TXG_INITIAL) {
1846 		metaslab_sync_done(ms, 0);
1847 		metaslab_space_update(vd, mg->mg_class,
1848 		    metaslab_allocated_space(ms), 0, 0);
1849 	}
1850 
1851 	/*
1852 	 * If metaslab_debug_load is set and we're initializing a metaslab
1853 	 * that has an allocated space map object then load the space map
1854 	 * so that we can verify frees.
1855 	 */
1856 	if (metaslab_debug_load && ms->ms_sm != NULL) {
1857 		mutex_enter(&ms->ms_lock);
1858 		VERIFY0(metaslab_load(ms));
1859 		mutex_exit(&ms->ms_lock);
1860 	}
1861 
1862 	if (txg != 0) {
1863 		vdev_dirty(vd, 0, NULL, txg);
1864 		vdev_dirty(vd, VDD_METASLAB, ms, txg);
1865 	}
1866 
1867 	*msp = ms;
1868 
1869 	return (0);
1870 }
1871 
1872 void
1873 metaslab_fini(metaslab_t *msp)
1874 {
1875 	metaslab_group_t *mg = msp->ms_group;
1876 	vdev_t *vd = mg->mg_vd;
1877 
1878 	metaslab_group_remove(mg, msp);
1879 
1880 	mutex_enter(&msp->ms_lock);
1881 	VERIFY(msp->ms_group == NULL);
1882 	metaslab_space_update(vd, mg->mg_class,
1883 	    -metaslab_allocated_space(msp), 0, -msp->ms_size);
1884 
1885 	space_map_close(msp->ms_sm);
1886 
1887 	metaslab_unload(msp);
1888 
1889 	range_tree_destroy(msp->ms_allocatable);
1890 	range_tree_destroy(msp->ms_freeing);
1891 	range_tree_destroy(msp->ms_freed);
1892 
1893 	for (int t = 0; t < TXG_SIZE; t++) {
1894 		range_tree_destroy(msp->ms_allocating[t]);
1895 	}
1896 
1897 	for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1898 		range_tree_destroy(msp->ms_defer[t]);
1899 	}
1900 	ASSERT0(msp->ms_deferspace);
1901 
1902 	range_tree_destroy(msp->ms_checkpointing);
1903 
1904 	for (int t = 0; t < TXG_SIZE; t++)
1905 		ASSERT(!txg_list_member(&vd->vdev_ms_list, msp, t));
1906 
1907 	range_tree_vacate(msp->ms_trim, NULL, NULL);
1908 	range_tree_destroy(msp->ms_trim);
1909 
1910 	mutex_exit(&msp->ms_lock);
1911 	cv_destroy(&msp->ms_load_cv);
1912 	mutex_destroy(&msp->ms_lock);
1913 	mutex_destroy(&msp->ms_sync_lock);
1914 	ASSERT3U(msp->ms_allocator, ==, -1);
1915 
1916 	kmem_free(msp, sizeof (metaslab_t));
1917 }
1918 
1919 #define	FRAGMENTATION_TABLE_SIZE	17
1920 
1921 /*
1922  * This table defines a segment size based fragmentation metric that will
1923  * allow each metaslab to derive its own fragmentation value. This is done
1924  * by calculating the space in each bucket of the spacemap histogram and
1925  * multiplying that by the fragmentation metric in this table. Doing
1926  * this for all buckets and dividing it by the total amount of free
1927  * space in this metaslab (i.e. the total free space in all buckets) gives
1928  * us the fragmentation metric. This means that a high fragmentation metric
1929  * equates to most of the free space being comprised of small segments.
1930  * Conversely, if the metric is low, then most of the free space is in
1931  * large segments. A 10% change in fragmentation equates to approximately
1932  * double the number of segments.
1933  *
1934  * This table defines 0% fragmented space using 16MB segments. Testing has
1935  * shown that segments that are greater than or equal to 16MB do not suffer
1936  * from drastic performance problems. Using this value, we derive the rest
1937  * of the table. Since the fragmentation value is never stored on disk, it
1938  * is possible to change these calculations in the future.
1939  */
1940 int zfs_frag_table[FRAGMENTATION_TABLE_SIZE] = {
1941 	100,	/* 512B	*/
1942 	100,	/* 1K	*/
1943 	98,	/* 2K	*/
1944 	95,	/* 4K	*/
1945 	90,	/* 8K	*/
1946 	80,	/* 16K	*/
1947 	70,	/* 32K	*/
1948 	60,	/* 64K	*/
1949 	50,	/* 128K	*/
1950 	40,	/* 256K	*/
1951 	30,	/* 512K	*/
1952 	20,	/* 1M	*/
1953 	15,	/* 2M	*/
1954 	10,	/* 4M	*/
1955 	5,	/* 8M	*/
1956 	0	/* 16M	*/
1957 };
1958 
1959 /*
1960  * Calculate the metaslab's fragmentation metric and set ms_fragmentation.
1961  * Setting this value to ZFS_FRAG_INVALID means that the metaslab has not
1962  * been upgraded and does not support this metric. Otherwise, the return
1963  * value should be in the range [0, 100].
1964  */
1965 static void
1966 metaslab_set_fragmentation(metaslab_t *msp)
1967 {
1968 	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1969 	uint64_t fragmentation = 0;
1970 	uint64_t total = 0;
1971 	boolean_t feature_enabled = spa_feature_is_enabled(spa,
1972 	    SPA_FEATURE_SPACEMAP_HISTOGRAM);
1973 
1974 	if (!feature_enabled) {
1975 		msp->ms_fragmentation = ZFS_FRAG_INVALID;
1976 		return;
1977 	}
1978 
1979 	/*
1980 	 * A null space map means that the entire metaslab is free
1981 	 * and thus is not fragmented.
1982 	 */
1983 	if (msp->ms_sm == NULL) {
1984 		msp->ms_fragmentation = 0;
1985 		return;
1986 	}
1987 
1988 	/*
1989 	 * If this metaslab's space map has not been upgraded, flag it
1990 	 * so that we upgrade next time we encounter it.
1991 	 */
1992 	if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t)) {
1993 		uint64_t txg = spa_syncing_txg(spa);
1994 		vdev_t *vd = msp->ms_group->mg_vd;
1995 
1996 		/*
1997 		 * If we've reached the final dirty txg, then we must
1998 		 * be shutting down the pool. We don't want to dirty
1999 		 * any data past this point so skip setting the condense
2000 		 * flag. We can retry this action the next time the pool
2001 		 * is imported.
2002 		 */
2003 		if (spa_writeable(spa) && txg < spa_final_dirty_txg(spa)) {
2004 			msp->ms_condense_wanted = B_TRUE;
2005 			vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
2006 			zfs_dbgmsg("txg %llu, requesting force condense: "
2007 			    "ms_id %llu, vdev_id %llu", txg, msp->ms_id,
2008 			    vd->vdev_id);
2009 		}
2010 		msp->ms_fragmentation = ZFS_FRAG_INVALID;
2011 		return;
2012 	}
2013 
2014 	for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
2015 		uint64_t space = 0;
2016 		uint8_t shift = msp->ms_sm->sm_shift;
2017 
2018 		int idx = MIN(shift - SPA_MINBLOCKSHIFT + i,
2019 		    FRAGMENTATION_TABLE_SIZE - 1);
2020 
2021 		if (msp->ms_sm->sm_phys->smp_histogram[i] == 0)
2022 			continue;
2023 
2024 		space = msp->ms_sm->sm_phys->smp_histogram[i] << (i + shift);
2025 		total += space;
2026 
2027 		ASSERT3U(idx, <, FRAGMENTATION_TABLE_SIZE);
2028 		fragmentation += space * zfs_frag_table[idx];
2029 	}
2030 
2031 	if (total > 0)
2032 		fragmentation /= total;
2033 	ASSERT3U(fragmentation, <=, 100);
2034 
2035 	msp->ms_fragmentation = fragmentation;
2036 }
2037 
2038 /*
2039  * Compute a weight -- a selection preference value -- for the given metaslab.
2040  * This is based on the amount of free space, the level of fragmentation,
2041  * the LBA range, and whether the metaslab is loaded.
2042  */
2043 static uint64_t
2044 metaslab_space_weight(metaslab_t *msp)
2045 {
2046 	metaslab_group_t *mg = msp->ms_group;
2047 	vdev_t *vd = mg->mg_vd;
2048 	uint64_t weight, space;
2049 
2050 	ASSERT(MUTEX_HELD(&msp->ms_lock));
2051 	ASSERT(!vd->vdev_removing);
2052 
2053 	/*
2054 	 * The baseline weight is the metaslab's free space.
2055 	 */
2056 	space = msp->ms_size - metaslab_allocated_space(msp);
2057 
2058 	if (metaslab_fragmentation_factor_enabled &&
2059 	    msp->ms_fragmentation != ZFS_FRAG_INVALID) {
2060 		/*
2061 		 * Use the fragmentation information to inversely scale
2062 		 * down the baseline weight. We need to ensure that we
2063 		 * don't exclude this metaslab completely when it's 100%
2064 		 * fragmented. To avoid this we reduce the fragmented value
2065 		 * by 1.
2066 		 */
2067 		space = (space * (100 - (msp->ms_fragmentation - 1))) / 100;
2068 
2069 		/*
2070 		 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
2071 		 * this metaslab again. The fragmentation metric may have
2072 		 * decreased the space to something smaller than
2073 		 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
2074 		 * so that we can consume any remaining space.
2075 		 */
2076 		if (space > 0 && space < SPA_MINBLOCKSIZE)
2077 			space = SPA_MINBLOCKSIZE;
2078 	}
2079 	weight = space;
2080 
2081 	/*
2082 	 * Modern disks have uniform bit density and constant angular velocity.
2083 	 * Therefore, the outer recording zones are faster (higher bandwidth)
2084 	 * than the inner zones by the ratio of outer to inner track diameter,
2085 	 * which is typically around 2:1.  We account for this by assigning
2086 	 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
2087 	 * In effect, this means that we'll select the metaslab with the most
2088 	 * free bandwidth rather than simply the one with the most free space.
2089 	 */
2090 	if (!vd->vdev_nonrot && metaslab_lba_weighting_enabled) {
2091 		weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count;
2092 		ASSERT(weight >= space && weight <= 2 * space);
2093 	}
2094 
2095 	/*
2096 	 * If this metaslab is one we're actively using, adjust its
2097 	 * weight to make it preferable to any inactive metaslab so
2098 	 * we'll polish it off. If the fragmentation on this metaslab
2099 	 * has exceed our threshold, then don't mark it active.
2100 	 */
2101 	if (msp->ms_loaded && msp->ms_fragmentation != ZFS_FRAG_INVALID &&
2102 	    msp->ms_fragmentation <= zfs_metaslab_fragmentation_threshold) {
2103 		weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
2104 	}
2105 
2106 	WEIGHT_SET_SPACEBASED(weight);
2107 	return (weight);
2108 }
2109 
2110 /*
2111  * Return the weight of the specified metaslab, according to the segment-based
2112  * weighting algorithm. The metaslab must be loaded. This function can
2113  * be called within a sync pass since it relies only on the metaslab's
2114  * range tree which is always accurate when the metaslab is loaded.
2115  */
2116 static uint64_t
2117 metaslab_weight_from_range_tree(metaslab_t *msp)
2118 {
2119 	uint64_t weight = 0;
2120 	uint32_t segments = 0;
2121 
2122 	ASSERT(msp->ms_loaded);
2123 
2124 	for (int i = RANGE_TREE_HISTOGRAM_SIZE - 1; i >= SPA_MINBLOCKSHIFT;
2125 	    i--) {
2126 		uint8_t shift = msp->ms_group->mg_vd->vdev_ashift;
2127 		int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1;
2128 
2129 		segments <<= 1;
2130 		segments += msp->ms_allocatable->rt_histogram[i];
2131 
2132 		/*
2133 		 * The range tree provides more precision than the space map
2134 		 * and must be downgraded so that all values fit within the
2135 		 * space map's histogram. This allows us to compare loaded
2136 		 * vs. unloaded metaslabs to determine which metaslab is
2137 		 * considered "best".
2138 		 */
2139 		if (i > max_idx)
2140 			continue;
2141 
2142 		if (segments != 0) {
2143 			WEIGHT_SET_COUNT(weight, segments);
2144 			WEIGHT_SET_INDEX(weight, i);
2145 			WEIGHT_SET_ACTIVE(weight, 0);
2146 			break;
2147 		}
2148 	}
2149 	return (weight);
2150 }
2151 
2152 /*
2153  * Calculate the weight based on the on-disk histogram. This should only
2154  * be called after a sync pass has completely finished since the on-disk
2155  * information is updated in metaslab_sync().
2156  */
2157 static uint64_t
2158 metaslab_weight_from_spacemap(metaslab_t *msp)
2159 {
2160 	space_map_t *sm = msp->ms_sm;
2161 	ASSERT(!msp->ms_loaded);
2162 	ASSERT(sm != NULL);
2163 	ASSERT3U(space_map_object(sm), !=, 0);
2164 	ASSERT3U(sm->sm_dbuf->db_size, ==, sizeof (space_map_phys_t));
2165 
2166 	/*
2167 	 * Create a joint histogram from all the segments that have made
2168 	 * it to the metaslab's space map histogram, that are not yet
2169 	 * available for allocation because they are still in the freeing
2170 	 * pipeline (e.g. freeing, freed, and defer trees). Then subtract
2171 	 * these segments from the space map's histogram to get a more
2172 	 * accurate weight.
2173 	 */
2174 	uint64_t deferspace_histogram[SPACE_MAP_HISTOGRAM_SIZE] = {0};
2175 	for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++)
2176 		deferspace_histogram[i] += msp->ms_synchist[i];
2177 	for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2178 		for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
2179 			deferspace_histogram[i] += msp->ms_deferhist[t][i];
2180 		}
2181 	}
2182 
2183 	uint64_t weight = 0;
2184 	for (int i = SPACE_MAP_HISTOGRAM_SIZE - 1; i >= 0; i--) {
2185 		ASSERT3U(sm->sm_phys->smp_histogram[i], >=,
2186 		    deferspace_histogram[i]);
2187 		uint64_t count =
2188 		    sm->sm_phys->smp_histogram[i] - deferspace_histogram[i];
2189 		if (count != 0) {
2190 			WEIGHT_SET_COUNT(weight, count);
2191 			WEIGHT_SET_INDEX(weight, i + sm->sm_shift);
2192 			WEIGHT_SET_ACTIVE(weight, 0);
2193 			break;
2194 		}
2195 	}
2196 	return (weight);
2197 }
2198 
2199 /*
2200  * Compute a segment-based weight for the specified metaslab. The weight
2201  * is determined by highest bucket in the histogram. The information
2202  * for the highest bucket is encoded into the weight value.
2203  */
2204 static uint64_t
2205 metaslab_segment_weight(metaslab_t *msp)
2206 {
2207 	metaslab_group_t *mg = msp->ms_group;
2208 	uint64_t weight = 0;
2209 	uint8_t shift = mg->mg_vd->vdev_ashift;
2210 
2211 	ASSERT(MUTEX_HELD(&msp->ms_lock));
2212 
2213 	/*
2214 	 * The metaslab is completely free.
2215 	 */
2216 	if (metaslab_allocated_space(msp) == 0) {
2217 		int idx = highbit64(msp->ms_size) - 1;
2218 		int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1;
2219 
2220 		if (idx < max_idx) {
2221 			WEIGHT_SET_COUNT(weight, 1ULL);
2222 			WEIGHT_SET_INDEX(weight, idx);
2223 		} else {
2224 			WEIGHT_SET_COUNT(weight, 1ULL << (idx - max_idx));
2225 			WEIGHT_SET_INDEX(weight, max_idx);
2226 		}
2227 		WEIGHT_SET_ACTIVE(weight, 0);
2228 		ASSERT(!WEIGHT_IS_SPACEBASED(weight));
2229 
2230 		return (weight);
2231 	}
2232 
2233 	ASSERT3U(msp->ms_sm->sm_dbuf->db_size, ==, sizeof (space_map_phys_t));
2234 
2235 	/*
2236 	 * If the metaslab is fully allocated then just make the weight 0.
2237 	 */
2238 	if (metaslab_allocated_space(msp) == msp->ms_size)
2239 		return (0);
2240 	/*
2241 	 * If the metaslab is already loaded, then use the range tree to
2242 	 * determine the weight. Otherwise, we rely on the space map information
2243 	 * to generate the weight.
2244 	 */
2245 	if (msp->ms_loaded) {
2246 		weight = metaslab_weight_from_range_tree(msp);
2247 	} else {
2248 		weight = metaslab_weight_from_spacemap(msp);
2249 	}
2250 
2251 	/*
2252 	 * If the metaslab was active the last time we calculated its weight
2253 	 * then keep it active. We want to consume the entire region that
2254 	 * is associated with this weight.
2255 	 */
2256 	if (msp->ms_activation_weight != 0 && weight != 0)
2257 		WEIGHT_SET_ACTIVE(weight, WEIGHT_GET_ACTIVE(msp->ms_weight));
2258 	return (weight);
2259 }
2260 
2261 /*
2262  * Determine if we should attempt to allocate from this metaslab. If the
2263  * metaslab has a maximum size then we can quickly determine if the desired
2264  * allocation size can be satisfied. Otherwise, if we're using segment-based
2265  * weighting then we can determine the maximum allocation that this metaslab
2266  * can accommodate based on the index encoded in the weight. If we're using
2267  * space-based weights then rely on the entire weight (excluding the weight
2268  * type bit).
2269  */
2270 boolean_t
2271 metaslab_should_allocate(metaslab_t *msp, uint64_t asize)
2272 {
2273 	boolean_t should_allocate;
2274 
2275 	if (msp->ms_max_size != 0)
2276 		return (msp->ms_max_size >= asize);
2277 
2278 	if (!WEIGHT_IS_SPACEBASED(msp->ms_weight)) {
2279 		/*
2280 		 * The metaslab segment weight indicates segments in the
2281 		 * range [2^i, 2^(i+1)), where i is the index in the weight.
2282 		 * Since the asize might be in the middle of the range, we
2283 		 * should attempt the allocation if asize < 2^(i+1).
2284 		 */
2285 		should_allocate = (asize <
2286 		    1ULL << (WEIGHT_GET_INDEX(msp->ms_weight) + 1));
2287 	} else {
2288 		should_allocate = (asize <=
2289 		    (msp->ms_weight & ~METASLAB_WEIGHT_TYPE));
2290 	}
2291 	return (should_allocate);
2292 }
2293 
2294 static uint64_t
2295 metaslab_weight(metaslab_t *msp)
2296 {
2297 	vdev_t *vd = msp->ms_group->mg_vd;
2298 	spa_t *spa = vd->vdev_spa;
2299 	uint64_t weight;
2300 
2301 	ASSERT(MUTEX_HELD(&msp->ms_lock));
2302 
2303 	/*
2304 	 * If this vdev is in the process of being removed, there is nothing
2305 	 * for us to do here.
2306 	 */
2307 	if (vd->vdev_removing)
2308 		return (0);
2309 
2310 	metaslab_set_fragmentation(msp);
2311 
2312 	/*
2313 	 * Update the maximum size if the metaslab is loaded. This will
2314 	 * ensure that we get an accurate maximum size if newly freed space
2315 	 * has been added back into the free tree.
2316 	 */
2317 	if (msp->ms_loaded)
2318 		msp->ms_max_size = metaslab_block_maxsize(msp);
2319 	else
2320 		ASSERT0(msp->ms_max_size);
2321 
2322 	/*
2323 	 * Segment-based weighting requires space map histogram support.
2324 	 */
2325 	if (zfs_metaslab_segment_weight_enabled &&
2326 	    spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
2327 	    (msp->ms_sm == NULL || msp->ms_sm->sm_dbuf->db_size ==
2328 	    sizeof (space_map_phys_t))) {
2329 		weight = metaslab_segment_weight(msp);
2330 	} else {
2331 		weight = metaslab_space_weight(msp);
2332 	}
2333 	return (weight);
2334 }
2335 
2336 void
2337 metaslab_recalculate_weight_and_sort(metaslab_t *msp)
2338 {
2339 	/* note: we preserve the mask (e.g. indication of primary, etc..) */
2340 	uint64_t was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
2341 	metaslab_group_sort(msp->ms_group, msp,
2342 	    metaslab_weight(msp) | was_active);
2343 }
2344 
2345 static int
2346 metaslab_activate_allocator(metaslab_group_t *mg, metaslab_t *msp,
2347     int allocator, uint64_t activation_weight)
2348 {
2349 	/*
2350 	 * If we're activating for the claim code, we don't want to actually
2351 	 * set the metaslab up for a specific allocator.
2352 	 */
2353 	if (activation_weight == METASLAB_WEIGHT_CLAIM)
2354 		return (0);
2355 	metaslab_t **arr = (activation_weight == METASLAB_WEIGHT_PRIMARY ?
2356 	    mg->mg_primaries : mg->mg_secondaries);
2357 
2358 	ASSERT(MUTEX_HELD(&msp->ms_lock));
2359 	mutex_enter(&mg->mg_lock);
2360 	if (arr[allocator] != NULL) {
2361 		mutex_exit(&mg->mg_lock);
2362 		return (EEXIST);
2363 	}
2364 
2365 	arr[allocator] = msp;
2366 	ASSERT3S(msp->ms_allocator, ==, -1);
2367 	msp->ms_allocator = allocator;
2368 	msp->ms_primary = (activation_weight == METASLAB_WEIGHT_PRIMARY);
2369 	mutex_exit(&mg->mg_lock);
2370 
2371 	return (0);
2372 }
2373 
2374 static int
2375 metaslab_activate(metaslab_t *msp, int allocator, uint64_t activation_weight)
2376 {
2377 	ASSERT(MUTEX_HELD(&msp->ms_lock));
2378 
2379 	if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
2380 		int error = metaslab_load(msp);
2381 		if (error != 0) {
2382 			metaslab_group_sort(msp->ms_group, msp, 0);
2383 			return (error);
2384 		}
2385 		if ((msp->ms_weight & METASLAB_ACTIVE_MASK) != 0) {
2386 			/*
2387 			 * The metaslab was activated for another allocator
2388 			 * while we were waiting, we should reselect.
2389 			 */
2390 			return (EBUSY);
2391 		}
2392 		if ((error = metaslab_activate_allocator(msp->ms_group, msp,
2393 		    allocator, activation_weight)) != 0) {
2394 			return (error);
2395 		}
2396 
2397 		msp->ms_activation_weight = msp->ms_weight;
2398 		metaslab_group_sort(msp->ms_group, msp,
2399 		    msp->ms_weight | activation_weight);
2400 	}
2401 	ASSERT(msp->ms_loaded);
2402 	ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
2403 
2404 	return (0);
2405 }
2406 
2407 static void
2408 metaslab_passivate_allocator(metaslab_group_t *mg, metaslab_t *msp,
2409     uint64_t weight)
2410 {
2411 	ASSERT(MUTEX_HELD(&msp->ms_lock));
2412 	if (msp->ms_weight & METASLAB_WEIGHT_CLAIM) {
2413 		metaslab_group_sort(mg, msp, weight);
2414 		return;
2415 	}
2416 
2417 	mutex_enter(&mg->mg_lock);
2418 	ASSERT3P(msp->ms_group, ==, mg);
2419 	if (msp->ms_primary) {
2420 		ASSERT3U(0, <=, msp->ms_allocator);
2421 		ASSERT3U(msp->ms_allocator, <, mg->mg_allocators);
2422 		ASSERT3P(mg->mg_primaries[msp->ms_allocator], ==, msp);
2423 		ASSERT(msp->ms_weight & METASLAB_WEIGHT_PRIMARY);
2424 		mg->mg_primaries[msp->ms_allocator] = NULL;
2425 	} else {
2426 		ASSERT(msp->ms_weight & METASLAB_WEIGHT_SECONDARY);
2427 		ASSERT3P(mg->mg_secondaries[msp->ms_allocator], ==, msp);
2428 		mg->mg_secondaries[msp->ms_allocator] = NULL;
2429 	}
2430 	msp->ms_allocator = -1;
2431 	metaslab_group_sort_impl(mg, msp, weight);
2432 	mutex_exit(&mg->mg_lock);
2433 }
2434 
2435 static void
2436 metaslab_passivate(metaslab_t *msp, uint64_t weight)
2437 {
2438 	uint64_t size = weight & ~METASLAB_WEIGHT_TYPE;
2439 
2440 	/*
2441 	 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
2442 	 * this metaslab again.  In that case, it had better be empty,
2443 	 * or we would be leaving space on the table.
2444 	 */
2445 	ASSERT(size >= SPA_MINBLOCKSIZE ||
2446 	    range_tree_is_empty(msp->ms_allocatable));
2447 	ASSERT0(weight & METASLAB_ACTIVE_MASK);
2448 
2449 	msp->ms_activation_weight = 0;
2450 	metaslab_passivate_allocator(msp->ms_group, msp, weight);
2451 	ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
2452 }
2453 
2454 /*
2455  * Segment-based metaslabs are activated once and remain active until
2456  * we either fail an allocation attempt (similar to space-based metaslabs)
2457  * or have exhausted the free space in zfs_metaslab_switch_threshold
2458  * buckets since the metaslab was activated. This function checks to see
2459  * if we've exhaused the zfs_metaslab_switch_threshold buckets in the
2460  * metaslab and passivates it proactively. This will allow us to select a
2461  * metaslabs with larger contiguous region if any remaining within this
2462  * metaslab group. If we're in sync pass > 1, then we continue using this
2463  * metaslab so that we don't dirty more block and cause more sync passes.
2464  */
2465 void
2466 metaslab_segment_may_passivate(metaslab_t *msp)
2467 {
2468 	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2469 
2470 	if (WEIGHT_IS_SPACEBASED(msp->ms_weight) || spa_sync_pass(spa) > 1)
2471 		return;
2472 
2473 	/*
2474 	 * Since we are in the middle of a sync pass, the most accurate
2475 	 * information that is accessible to us is the in-core range tree
2476 	 * histogram; calculate the new weight based on that information.
2477 	 */
2478 	uint64_t weight = metaslab_weight_from_range_tree(msp);
2479 	int activation_idx = WEIGHT_GET_INDEX(msp->ms_activation_weight);
2480 	int current_idx = WEIGHT_GET_INDEX(weight);
2481 
2482 	if (current_idx <= activation_idx - zfs_metaslab_switch_threshold)
2483 		metaslab_passivate(msp, weight);
2484 }
2485 
2486 static void
2487 metaslab_preload(void *arg)
2488 {
2489 	metaslab_t *msp = arg;
2490 	spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2491 
2492 	ASSERT(!MUTEX_HELD(&msp->ms_group->mg_lock));
2493 
2494 	mutex_enter(&msp->ms_lock);
2495 	(void) metaslab_load(msp);
2496 	msp->ms_selected_txg = spa_syncing_txg(spa);
2497 	mutex_exit(&msp->ms_lock);
2498 }
2499 
2500 static void
2501 metaslab_group_preload(metaslab_group_t *mg)
2502 {
2503 	spa_t *spa = mg->mg_vd->vdev_spa;
2504 	metaslab_t *msp;
2505 	avl_tree_t *t = &mg->mg_metaslab_tree;
2506 	int m = 0;
2507 
2508 	if (spa_shutting_down(spa) || !metaslab_preload_enabled) {
2509 		taskq_wait(mg->mg_taskq);
2510 		return;
2511 	}
2512 
2513 	mutex_enter(&mg->mg_lock);
2514 
2515 	/*
2516 	 * Load the next potential metaslabs
2517 	 */
2518 	for (msp = avl_first(t); msp != NULL; msp = AVL_NEXT(t, msp)) {
2519 		ASSERT3P(msp->ms_group, ==, mg);
2520 
2521 		/*
2522 		 * We preload only the maximum number of metaslabs specified
2523 		 * by metaslab_preload_limit. If a metaslab is being forced
2524 		 * to condense then we preload it too. This will ensure
2525 		 * that force condensing happens in the next txg.
2526 		 */
2527 		if (++m > metaslab_preload_limit && !msp->ms_condense_wanted) {
2528 			continue;
2529 		}
2530 
2531 		VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
2532 		    msp, TQ_SLEEP) != TASKQID_INVALID);
2533 	}
2534 	mutex_exit(&mg->mg_lock);
2535 }
2536 
2537 /*
2538  * Determine if the space map's on-disk footprint is past our tolerance
2539  * for inefficiency. We would like to use the following criteria to make
2540  * our decision:
2541  *
2542  * 1. The size of the space map object should not dramatically increase as a
2543  * result of writing out the free space range tree.
2544  *
2545  * 2. The minimal on-disk space map representation is zfs_condense_pct/100
2546  * times the size than the free space range tree representation
2547  * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1MB).
2548  *
2549  * 3. The on-disk size of the space map should actually decrease.
2550  *
2551  * Unfortunately, we cannot compute the on-disk size of the space map in this
2552  * context because we cannot accurately compute the effects of compression, etc.
2553  * Instead, we apply the heuristic described in the block comment for
2554  * zfs_metaslab_condense_block_threshold - we only condense if the space used
2555  * is greater than a threshold number of blocks.
2556  */
2557 static boolean_t
2558 metaslab_should_condense(metaslab_t *msp)
2559 {
2560 	space_map_t *sm = msp->ms_sm;
2561 	vdev_t *vd = msp->ms_group->mg_vd;
2562 	uint64_t vdev_blocksize = 1 << vd->vdev_ashift;
2563 	uint64_t current_txg = spa_syncing_txg(vd->vdev_spa);
2564 
2565 	ASSERT(MUTEX_HELD(&msp->ms_lock));
2566 	ASSERT(msp->ms_loaded);
2567 
2568 	/*
2569 	 * Allocations and frees in early passes are generally more space
2570 	 * efficient (in terms of blocks described in space map entries)
2571 	 * than the ones in later passes (e.g. we don't compress after
2572 	 * sync pass 5) and condensing a metaslab multiple times in a txg
2573 	 * could degrade performance.
2574 	 *
2575 	 * Thus we prefer condensing each metaslab at most once every txg at
2576 	 * the earliest sync pass possible. If a metaslab is eligible for
2577 	 * condensing again after being considered for condensing within the
2578 	 * same txg, it will hopefully be dirty in the next txg where it will
2579 	 * be condensed at an earlier pass.
2580 	 */
2581 	if (msp->ms_condense_checked_txg == current_txg)
2582 		return (B_FALSE);
2583 	msp->ms_condense_checked_txg = current_txg;
2584 
2585 	/*
2586 	 * We always condense metaslabs that are empty and metaslabs for
2587 	 * which a condense request has been made.
2588 	 */
2589 	if (avl_is_empty(&msp->ms_allocatable_by_size) ||
2590 	    msp->ms_condense_wanted)
2591 		return (B_TRUE);
2592 
2593 	uint64_t object_size = space_map_length(msp->ms_sm);
2594 	uint64_t optimal_size = space_map_estimate_optimal_size(sm,
2595 	    msp->ms_allocatable, SM_NO_VDEVID);
2596 
2597 	dmu_object_info_t doi;
2598 	dmu_object_info_from_db(sm->sm_dbuf, &doi);
2599 	uint64_t record_size = MAX(doi.doi_data_block_size, vdev_blocksize);
2600 
2601 	return (object_size >= (optimal_size * zfs_condense_pct / 100) &&
2602 	    object_size > zfs_metaslab_condense_block_threshold * record_size);
2603 }
2604 
2605 /*
2606  * Condense the on-disk space map representation to its minimized form.
2607  * The minimized form consists of a small number of allocations followed by
2608  * the entries of the free range tree.
2609  */
2610 static void
2611 metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
2612 {
2613 	range_tree_t *condense_tree;
2614 	space_map_t *sm = msp->ms_sm;
2615 
2616 	ASSERT(MUTEX_HELD(&msp->ms_lock));
2617 	ASSERT(msp->ms_loaded);
2618 
2619 	zfs_dbgmsg("condensing: txg %llu, msp[%llu] %p, vdev id %llu, "
2620 	    "spa %s, smp size %llu, segments %lu, forcing condense=%s", txg,
2621 	    msp->ms_id, msp, msp->ms_group->mg_vd->vdev_id,
2622 	    msp->ms_group->mg_vd->vdev_spa->spa_name,
2623 	    space_map_length(msp->ms_sm),
2624 	    avl_numnodes(&msp->ms_allocatable->rt_root),
2625 	    msp->ms_condense_wanted ? "TRUE" : "FALSE");
2626 
2627 	msp->ms_condense_wanted = B_FALSE;
2628 
2629 	/*
2630 	 * Create an range tree that is 100% allocated. We remove segments
2631 	 * that have been freed in this txg, any deferred frees that exist,
2632 	 * and any allocation in the future. Removing segments should be
2633 	 * a relatively inexpensive operation since we expect these trees to
2634 	 * have a small number of nodes.
2635 	 */
2636 	condense_tree = range_tree_create(NULL, NULL);
2637 	range_tree_add(condense_tree, msp->ms_start, msp->ms_size);
2638 
2639 	range_tree_walk(msp->ms_freeing, range_tree_remove, condense_tree);
2640 	range_tree_walk(msp->ms_freed, range_tree_remove, condense_tree);
2641 
2642 	for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2643 		range_tree_walk(msp->ms_defer[t],
2644 		    range_tree_remove, condense_tree);
2645 	}
2646 
2647 	for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
2648 		range_tree_walk(msp->ms_allocating[(txg + t) & TXG_MASK],
2649 		    range_tree_remove, condense_tree);
2650 	}
2651 
2652 	/*
2653 	 * We're about to drop the metaslab's lock thus allowing
2654 	 * other consumers to change it's content. Set the
2655 	 * metaslab's ms_condensing flag to ensure that
2656 	 * allocations on this metaslab do not occur while we're
2657 	 * in the middle of committing it to disk. This is only critical
2658 	 * for ms_allocatable as all other range trees use per txg
2659 	 * views of their content.
2660 	 */
2661 	msp->ms_condensing = B_TRUE;
2662 
2663 	mutex_exit(&msp->ms_lock);
2664 	space_map_truncate(sm, zfs_metaslab_sm_blksz, tx);
2665 
2666 	/*
2667 	 * While we would ideally like to create a space map representation
2668 	 * that consists only of allocation records, doing so can be
2669 	 * prohibitively expensive because the in-core free tree can be
2670 	 * large, and therefore computationally expensive to subtract
2671 	 * from the condense_tree. Instead we sync out two trees, a cheap
2672 	 * allocation only tree followed by the in-core free tree. While not
2673 	 * optimal, this is typically close to optimal, and much cheaper to
2674 	 * compute.
2675 	 */
2676 	space_map_write(sm, condense_tree, SM_ALLOC, SM_NO_VDEVID, tx);
2677 	range_tree_vacate(condense_tree, NULL, NULL);
2678 	range_tree_destroy(condense_tree);
2679 
2680 	space_map_write(sm, msp->ms_allocatable, SM_FREE, SM_NO_VDEVID, tx);
2681 	mutex_enter(&msp->ms_lock);
2682 	msp->ms_condensing = B_FALSE;
2683 }
2684 
2685 /*
2686  * Write a metaslab to disk in the context of the specified transaction group.
2687  */
2688 void
2689 metaslab_sync(metaslab_t *msp, uint64_t txg)
2690 {
2691 	metaslab_group_t *mg = msp->ms_group;
2692 	vdev_t *vd = mg->mg_vd;
2693 	spa_t *spa = vd->vdev_spa;
2694 	objset_t *mos = spa_meta_objset(spa);
2695 	range_tree_t *alloctree = msp->ms_allocating[txg & TXG_MASK];
2696 	dmu_tx_t *tx;
2697 	uint64_t object = space_map_object(msp->ms_sm);
2698 
2699 	ASSERT(!vd->vdev_ishole);
2700 
2701 	/*
2702 	 * This metaslab has just been added so there's no work to do now.
2703 	 */
2704 	if (msp->ms_freeing == NULL) {
2705 		ASSERT3P(alloctree, ==, NULL);
2706 		return;
2707 	}
2708 
2709 	ASSERT3P(alloctree, !=, NULL);
2710 	ASSERT3P(msp->ms_freeing, !=, NULL);
2711 	ASSERT3P(msp->ms_freed, !=, NULL);
2712 	ASSERT3P(msp->ms_checkpointing, !=, NULL);
2713 	ASSERT3P(msp->ms_trim, !=, NULL);
2714 
2715 	/*
2716 	 * Normally, we don't want to process a metaslab if there are no
2717 	 * allocations or frees to perform. However, if the metaslab is being
2718 	 * forced to condense and it's loaded, we need to let it through.
2719 	 */
2720 	if (range_tree_is_empty(alloctree) &&
2721 	    range_tree_is_empty(msp->ms_freeing) &&
2722 	    range_tree_is_empty(msp->ms_checkpointing) &&
2723 	    !(msp->ms_loaded && msp->ms_condense_wanted))
2724 		return;
2725 
2726 
2727 	VERIFY(txg <= spa_final_dirty_txg(spa));
2728 
2729 	/*
2730 	 * The only state that can actually be changing concurrently
2731 	 * with metaslab_sync() is the metaslab's ms_allocatable. No
2732 	 * other thread can be modifying this txg's alloc, freeing,
2733 	 * freed, or space_map_phys_t.  We drop ms_lock whenever we
2734 	 * could call into the DMU, because the DMU can call down to
2735 	 * us (e.g. via zio_free()) at any time.
2736 	 *
2737 	 * The spa_vdev_remove_thread() can be reading metaslab state
2738 	 * concurrently, and it is locked out by the ms_sync_lock.
2739 	 * Note that the ms_lock is insufficient for this, because it
2740 	 * is dropped by space_map_write().
2741 	 */
2742 	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2743 
2744 	if (msp->ms_sm == NULL) {
2745 		uint64_t new_object;
2746 
2747 		new_object = space_map_alloc(mos, zfs_metaslab_sm_blksz, tx);
2748 		VERIFY3U(new_object, !=, 0);
2749 
2750 		VERIFY0(space_map_open(&msp->ms_sm, mos, new_object,
2751 		    msp->ms_start, msp->ms_size, vd->vdev_ashift));
2752 
2753 		ASSERT(msp->ms_sm != NULL);
2754 		ASSERT0(metaslab_allocated_space(msp));
2755 	}
2756 
2757 	if (!range_tree_is_empty(msp->ms_checkpointing) &&
2758 	    vd->vdev_checkpoint_sm == NULL) {
2759 		ASSERT(spa_has_checkpoint(spa));
2760 
2761 		uint64_t new_object = space_map_alloc(mos,
2762 		    vdev_standard_sm_blksz, tx);
2763 		VERIFY3U(new_object, !=, 0);
2764 
2765 		VERIFY0(space_map_open(&vd->vdev_checkpoint_sm,
2766 		    mos, new_object, 0, vd->vdev_asize, vd->vdev_ashift));
2767 		ASSERT3P(vd->vdev_checkpoint_sm, !=, NULL);
2768 
2769 		/*
2770 		 * We save the space map object as an entry in vdev_top_zap
2771 		 * so it can be retrieved when the pool is reopened after an
2772 		 * export or through zdb.
2773 		 */
2774 		VERIFY0(zap_add(vd->vdev_spa->spa_meta_objset,
2775 		    vd->vdev_top_zap, VDEV_TOP_ZAP_POOL_CHECKPOINT_SM,
2776 		    sizeof (new_object), 1, &new_object, tx));
2777 	}
2778 
2779 	mutex_enter(&msp->ms_sync_lock);
2780 	mutex_enter(&msp->ms_lock);
2781 
2782 	/*
2783 	 * Note: metaslab_condense() clears the space map's histogram.
2784 	 * Therefore we must verify and remove this histogram before
2785 	 * condensing.
2786 	 */
2787 	metaslab_group_histogram_verify(mg);
2788 	metaslab_class_histogram_verify(mg->mg_class);
2789 	metaslab_group_histogram_remove(mg, msp);
2790 
2791 	if (msp->ms_loaded && metaslab_should_condense(msp)) {
2792 		metaslab_condense(msp, txg, tx);
2793 	} else {
2794 		mutex_exit(&msp->ms_lock);
2795 		space_map_write(msp->ms_sm, alloctree, SM_ALLOC,
2796 		    SM_NO_VDEVID, tx);
2797 		space_map_write(msp->ms_sm, msp->ms_freeing, SM_FREE,
2798 		    SM_NO_VDEVID, tx);
2799 		mutex_enter(&msp->ms_lock);
2800 	}
2801 
2802 	msp->ms_allocated_space += range_tree_space(alloctree);
2803 	ASSERT3U(msp->ms_allocated_space, >=,
2804 	    range_tree_space(msp->ms_freeing));
2805 	msp->ms_allocated_space -= range_tree_space(msp->ms_freeing);
2806 
2807 	if (!range_tree_is_empty(msp->ms_checkpointing)) {
2808 		ASSERT(spa_has_checkpoint(spa));
2809 		ASSERT3P(vd->vdev_checkpoint_sm, !=, NULL);
2810 
2811 		/*
2812 		 * Since we are doing writes to disk and the ms_checkpointing
2813 		 * tree won't be changing during that time, we drop the
2814 		 * ms_lock while writing to the checkpoint space map.
2815 		 */
2816 		mutex_exit(&msp->ms_lock);
2817 		space_map_write(vd->vdev_checkpoint_sm,
2818 		    msp->ms_checkpointing, SM_FREE, SM_NO_VDEVID, tx);
2819 		mutex_enter(&msp->ms_lock);
2820 
2821 		spa->spa_checkpoint_info.sci_dspace +=
2822 		    range_tree_space(msp->ms_checkpointing);
2823 		vd->vdev_stat.vs_checkpoint_space +=
2824 		    range_tree_space(msp->ms_checkpointing);
2825 		ASSERT3U(vd->vdev_stat.vs_checkpoint_space, ==,
2826 		    -space_map_allocated(vd->vdev_checkpoint_sm));
2827 
2828 		range_tree_vacate(msp->ms_checkpointing, NULL, NULL);
2829 	}
2830 
2831 	if (msp->ms_loaded) {
2832 		/*
2833 		 * When the space map is loaded, we have an accurate
2834 		 * histogram in the range tree. This gives us an opportunity
2835 		 * to bring the space map's histogram up-to-date so we clear
2836 		 * it first before updating it.
2837 		 */
2838 		space_map_histogram_clear(msp->ms_sm);
2839 		space_map_histogram_add(msp->ms_sm, msp->ms_allocatable, tx);
2840 
2841 		/*
2842 		 * Since we've cleared the histogram we need to add back
2843 		 * any free space that has already been processed, plus
2844 		 * any deferred space. This allows the on-disk histogram
2845 		 * to accurately reflect all free space even if some space
2846 		 * is not yet available for allocation (i.e. deferred).
2847 		 */
2848 		space_map_histogram_add(msp->ms_sm, msp->ms_freed, tx);
2849 
2850 		/*
2851 		 * Add back any deferred free space that has not been
2852 		 * added back into the in-core free tree yet. This will
2853 		 * ensure that we don't end up with a space map histogram
2854 		 * that is completely empty unless the metaslab is fully
2855 		 * allocated.
2856 		 */
2857 		for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2858 			space_map_histogram_add(msp->ms_sm,
2859 			    msp->ms_defer[t], tx);
2860 		}
2861 	}
2862 
2863 	/*
2864 	 * Always add the free space from this sync pass to the space
2865 	 * map histogram. We want to make sure that the on-disk histogram
2866 	 * accounts for all free space. If the space map is not loaded,
2867 	 * then we will lose some accuracy but will correct it the next
2868 	 * time we load the space map.
2869 	 */
2870 	space_map_histogram_add(msp->ms_sm, msp->ms_freeing, tx);
2871 	metaslab_aux_histograms_update(msp);
2872 
2873 	metaslab_group_histogram_add(mg, msp);
2874 	metaslab_group_histogram_verify(mg);
2875 	metaslab_class_histogram_verify(mg->mg_class);
2876 
2877 	/*
2878 	 * For sync pass 1, we avoid traversing this txg's free range tree
2879 	 * and instead will just swap the pointers for freeing and freed.
2880 	 * We can safely do this since the freed_tree is guaranteed to be
2881 	 * empty on the initial pass.
2882 	 */
2883 	if (spa_sync_pass(spa) == 1) {
2884 		range_tree_swap(&msp->ms_freeing, &msp->ms_freed);
2885 		ASSERT0(msp->ms_allocated_this_txg);
2886 	} else {
2887 		range_tree_vacate(msp->ms_freeing,
2888 		    range_tree_add, msp->ms_freed);
2889 	}
2890 	msp->ms_allocated_this_txg += range_tree_space(alloctree);
2891 	range_tree_vacate(alloctree, NULL, NULL);
2892 
2893 	ASSERT0(range_tree_space(msp->ms_allocating[txg & TXG_MASK]));
2894 	ASSERT0(range_tree_space(msp->ms_allocating[TXG_CLEAN(txg)
2895 	    & TXG_MASK]));
2896 	ASSERT0(range_tree_space(msp->ms_freeing));
2897 	ASSERT0(range_tree_space(msp->ms_checkpointing));
2898 
2899 	mutex_exit(&msp->ms_lock);
2900 
2901 	if (object != space_map_object(msp->ms_sm)) {
2902 		object = space_map_object(msp->ms_sm);
2903 		dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
2904 		    msp->ms_id, sizeof (uint64_t), &object, tx);
2905 	}
2906 	mutex_exit(&msp->ms_sync_lock);
2907 	dmu_tx_commit(tx);
2908 }
2909 
2910 /*
2911  * Called after a transaction group has completely synced to mark
2912  * all of the metaslab's free space as usable.
2913  */
2914 void
2915 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
2916 {
2917 	metaslab_group_t *mg = msp->ms_group;
2918 	vdev_t *vd = mg->mg_vd;
2919 	spa_t *spa = vd->vdev_spa;
2920 	range_tree_t **defer_tree;
2921 	int64_t alloc_delta, defer_delta;
2922 	boolean_t defer_allowed = B_TRUE;
2923 
2924 	ASSERT(!vd->vdev_ishole);
2925 
2926 	mutex_enter(&msp->ms_lock);
2927 
2928 	/*
2929 	 * If this metaslab is just becoming available, initialize its
2930 	 * range trees and add its capacity to the vdev.
2931 	 */
2932 	if (msp->ms_freed == NULL) {
2933 		for (int t = 0; t < TXG_SIZE; t++) {
2934 			ASSERT(msp->ms_allocating[t] == NULL);
2935 
2936 			msp->ms_allocating[t] = range_tree_create(NULL, NULL);
2937 		}
2938 
2939 		ASSERT3P(msp->ms_freeing, ==, NULL);
2940 		msp->ms_freeing = range_tree_create(NULL, NULL);
2941 
2942 		ASSERT3P(msp->ms_freed, ==, NULL);
2943 		msp->ms_freed = range_tree_create(NULL, NULL);
2944 
2945 		for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2946 			ASSERT(msp->ms_defer[t] == NULL);
2947 
2948 			msp->ms_defer[t] = range_tree_create(NULL, NULL);
2949 		}
2950 
2951 		ASSERT3P(msp->ms_checkpointing, ==, NULL);
2952 		msp->ms_checkpointing = range_tree_create(NULL, NULL);
2953 
2954 		metaslab_space_update(vd, mg->mg_class, 0, 0, msp->ms_size);
2955 	}
2956 	ASSERT0(range_tree_space(msp->ms_freeing));
2957 	ASSERT0(range_tree_space(msp->ms_checkpointing));
2958 
2959 	defer_tree = &msp->ms_defer[txg % TXG_DEFER_SIZE];
2960 
2961 	uint64_t free_space = metaslab_class_get_space(spa_normal_class(spa)) -
2962 	    metaslab_class_get_alloc(spa_normal_class(spa));
2963 	if (free_space <= spa_get_slop_space(spa) || vd->vdev_removing) {
2964 		defer_allowed = B_FALSE;
2965 	}
2966 
2967 	defer_delta = 0;
2968 	alloc_delta = msp->ms_allocated_this_txg -
2969 	    range_tree_space(msp->ms_freed);
2970 	if (defer_allowed) {
2971 		defer_delta = range_tree_space(msp->ms_freed) -
2972 		    range_tree_space(*defer_tree);
2973 	} else {
2974 		defer_delta -= range_tree_space(*defer_tree);
2975 	}
2976 
2977 	metaslab_space_update(vd, mg->mg_class, alloc_delta + defer_delta,
2978 	    defer_delta, 0);
2979 
2980 	/*
2981 	 * If there's a metaslab_load() in progress, wait for it to complete
2982 	 * so that we have a consistent view of the in-core space map.
2983 	 */
2984 	metaslab_load_wait(msp);
2985 
2986 	/*
2987 	 * When auto-trimming is enabled, free ranges which are added to
2988 	 * ms_allocatable are also be added to ms_trim.  The ms_trim tree is
2989 	 * periodically consumed by the vdev_autotrim_thread() which issues
2990 	 * trims for all ranges and then vacates the tree.  The ms_trim tree
2991 	 * can be discarded at any time with the sole consequence of recent
2992 	 * frees not being trimmed.
2993 	 */
2994 	if (spa_get_autotrim(spa) == SPA_AUTOTRIM_ON) {
2995 		range_tree_walk(*defer_tree, range_tree_add, msp->ms_trim);
2996 		if (!defer_allowed) {
2997 			range_tree_walk(msp->ms_freed, range_tree_add,
2998 			    msp->ms_trim);
2999 		}
3000 	} else {
3001 		range_tree_vacate(msp->ms_trim, NULL, NULL);
3002 	}
3003 
3004 	/*
3005 	 * Move the frees from the defer_tree back to the free
3006 	 * range tree (if it's loaded). Swap the freed_tree and
3007 	 * the defer_tree -- this is safe to do because we've
3008 	 * just emptied out the defer_tree.
3009 	 */
3010 	range_tree_vacate(*defer_tree,
3011 	    msp->ms_loaded ? range_tree_add : NULL, msp->ms_allocatable);
3012 	if (defer_allowed) {
3013 		range_tree_swap(&msp->ms_freed, defer_tree);
3014 	} else {
3015 		range_tree_vacate(msp->ms_freed,
3016 		    msp->ms_loaded ? range_tree_add : NULL,
3017 		    msp->ms_allocatable);
3018 	}
3019 
3020 	msp->ms_synced_length = space_map_length(msp->ms_sm);
3021 
3022 	msp->ms_deferspace += defer_delta;
3023 	ASSERT3S(msp->ms_deferspace, >=, 0);
3024 	ASSERT3S(msp->ms_deferspace, <=, msp->ms_size);
3025 	if (msp->ms_deferspace != 0) {
3026 		/*
3027 		 * Keep syncing this metaslab until all deferred frees
3028 		 * are back in circulation.
3029 		 */
3030 		vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
3031 	}
3032 	metaslab_aux_histograms_update_done(msp, defer_allowed);
3033 
3034 	if (msp->ms_new) {
3035 		msp->ms_new = B_FALSE;
3036 		mutex_enter(&mg->mg_lock);
3037 		mg->mg_ms_ready++;
3038 		mutex_exit(&mg->mg_lock);
3039 	}
3040 
3041 	/*
3042 	 * Re-sort metaslab within its group now that we've adjusted
3043 	 * its allocatable space.
3044 	 */
3045 	metaslab_recalculate_weight_and_sort(msp);
3046 
3047 	/*
3048 	 * If the metaslab is loaded and we've not tried to load or allocate
3049 	 * from it in 'metaslab_unload_delay' txgs, then unload it.
3050 	 */
3051 	if (msp->ms_loaded &&
3052 	    msp->ms_disabled == 0 &&
3053 	    msp->ms_selected_txg + metaslab_unload_delay < txg) {
3054 
3055 		for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
3056 			VERIFY0(range_tree_space(
3057 			    msp->ms_allocating[(txg + t) & TXG_MASK]));
3058 		}
3059 		if (msp->ms_allocator != -1) {
3060 			metaslab_passivate(msp, msp->ms_weight &
3061 			    ~METASLAB_ACTIVE_MASK);
3062 		}
3063 
3064 		if (!metaslab_debug_unload)
3065 			metaslab_unload(msp);
3066 	}
3067 
3068 	ASSERT0(range_tree_space(msp->ms_allocating[txg & TXG_MASK]));
3069 	ASSERT0(range_tree_space(msp->ms_freeing));
3070 	ASSERT0(range_tree_space(msp->ms_freed));
3071 	ASSERT0(range_tree_space(msp->ms_checkpointing));
3072 
3073 	msp->ms_allocated_this_txg = 0;
3074 	mutex_exit(&msp->ms_lock);
3075 }
3076 
3077 void
3078 metaslab_sync_reassess(metaslab_group_t *mg)
3079 {
3080 	spa_t *spa = mg->mg_class->mc_spa;
3081 
3082 	spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
3083 	metaslab_group_alloc_update(mg);
3084 	mg->mg_fragmentation = metaslab_group_fragmentation(mg);
3085 
3086 	/*
3087 	 * Preload the next potential metaslabs but only on active
3088 	 * metaslab groups. We can get into a state where the metaslab
3089 	 * is no longer active since we dirty metaslabs as we remove a
3090 	 * a device, thus potentially making the metaslab group eligible
3091 	 * for preloading.
3092 	 */
3093 	if (mg->mg_activation_count > 0) {
3094 		metaslab_group_preload(mg);
3095 	}
3096 	spa_config_exit(spa, SCL_ALLOC, FTAG);
3097 }
3098 
3099 /*
3100  * When writing a ditto block (i.e. more than one DVA for a given BP) on
3101  * the same vdev as an existing DVA of this BP, then try to allocate it
3102  * on a different metaslab than existing DVAs (i.e. a unique metaslab).
3103  */
3104 static boolean_t
3105 metaslab_is_unique(metaslab_t *msp, dva_t *dva)
3106 {
3107 	uint64_t dva_ms_id;
3108 
3109 	if (DVA_GET_ASIZE(dva) == 0)
3110 		return (B_TRUE);
3111 
3112 	if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
3113 		return (B_TRUE);
3114 
3115 	dva_ms_id = DVA_GET_OFFSET(dva) >> msp->ms_group->mg_vd->vdev_ms_shift;
3116 
3117 	return (msp->ms_id != dva_ms_id);
3118 }
3119 
3120 /*
3121  * ==========================================================================
3122  * Metaslab allocation tracing facility
3123  * ==========================================================================
3124  */
3125 kstat_t *metaslab_trace_ksp;
3126 kstat_named_t metaslab_trace_over_limit;
3127 
3128 void
3129 metaslab_alloc_trace_init(void)
3130 {
3131 	ASSERT(metaslab_alloc_trace_cache == NULL);
3132 	metaslab_alloc_trace_cache = kmem_cache_create(
3133 	    "metaslab_alloc_trace_cache", sizeof (metaslab_alloc_trace_t),
3134 	    0, NULL, NULL, NULL, NULL, NULL, 0);
3135 	metaslab_trace_ksp = kstat_create("zfs", 0, "metaslab_trace_stats",
3136 	    "misc", KSTAT_TYPE_NAMED, 1, KSTAT_FLAG_VIRTUAL);
3137 	if (metaslab_trace_ksp != NULL) {
3138 		metaslab_trace_ksp->ks_data = &metaslab_trace_over_limit;
3139 		kstat_named_init(&metaslab_trace_over_limit,
3140 		    "metaslab_trace_over_limit", KSTAT_DATA_UINT64);
3141 		kstat_install(metaslab_trace_ksp);
3142 	}
3143 }
3144 
3145 void
3146 metaslab_alloc_trace_fini(void)
3147 {
3148 	if (metaslab_trace_ksp != NULL) {
3149 		kstat_delete(metaslab_trace_ksp);
3150 		metaslab_trace_ksp = NULL;
3151 	}
3152 	kmem_cache_destroy(metaslab_alloc_trace_cache);
3153 	metaslab_alloc_trace_cache = NULL;
3154 }
3155 
3156 /*
3157  * Add an allocation trace element to the allocation tracing list.
3158  */
3159 static void
3160 metaslab_trace_add(zio_alloc_list_t *zal, metaslab_group_t *mg,
3161     metaslab_t *msp, uint64_t psize, uint32_t dva_id, uint64_t offset,
3162     int allocator)
3163 {
3164 	if (!metaslab_trace_enabled)
3165 		return;
3166 
3167 	/*
3168 	 * When the tracing list reaches its maximum we remove
3169 	 * the second element in the list before adding a new one.
3170 	 * By removing the second element we preserve the original
3171 	 * entry as a clue to what allocations steps have already been
3172 	 * performed.
3173 	 */
3174 	if (zal->zal_size == metaslab_trace_max_entries) {
3175 		metaslab_alloc_trace_t *mat_next;
3176 #ifdef DEBUG
3177 		panic("too many entries in allocation list");
3178 #endif
3179 		atomic_inc_64(&metaslab_trace_over_limit.value.ui64);
3180 		zal->zal_size--;
3181 		mat_next = list_next(&zal->zal_list, list_head(&zal->zal_list));
3182 		list_remove(&zal->zal_list, mat_next);
3183 		kmem_cache_free(metaslab_alloc_trace_cache, mat_next);
3184 	}
3185 
3186 	metaslab_alloc_trace_t *mat =
3187 	    kmem_cache_alloc(metaslab_alloc_trace_cache, KM_SLEEP);
3188 	list_link_init(&mat->mat_list_node);
3189 	mat->mat_mg = mg;
3190 	mat->mat_msp = msp;
3191 	mat->mat_size = psize;
3192 	mat->mat_dva_id = dva_id;
3193 	mat->mat_offset = offset;
3194 	mat->mat_weight = 0;
3195 	mat->mat_allocator = allocator;
3196 
3197 	if (msp != NULL)
3198 		mat->mat_weight = msp->ms_weight;
3199 
3200 	/*
3201 	 * The list is part of the zio so locking is not required. Only
3202 	 * a single thread will perform allocations for a given zio.
3203 	 */
3204 	list_insert_tail(&zal->zal_list, mat);
3205 	zal->zal_size++;
3206 
3207 	ASSERT3U(zal->zal_size, <=, metaslab_trace_max_entries);
3208 }
3209 
3210 void
3211 metaslab_trace_init(zio_alloc_list_t *zal)
3212 {
3213 	list_create(&zal->zal_list, sizeof (metaslab_alloc_trace_t),
3214 	    offsetof(metaslab_alloc_trace_t, mat_list_node));
3215 	zal->zal_size = 0;
3216 }
3217 
3218 void
3219 metaslab_trace_fini(zio_alloc_list_t *zal)
3220 {
3221 	metaslab_alloc_trace_t *mat;
3222 
3223 	while ((mat = list_remove_head(&zal->zal_list)) != NULL)
3224 		kmem_cache_free(metaslab_alloc_trace_cache, mat);
3225 	list_destroy(&zal->zal_list);
3226 	zal->zal_size = 0;
3227 }
3228 
3229 /*
3230  * ==========================================================================
3231  * Metaslab block operations
3232  * ==========================================================================
3233  */
3234 
3235 static void
3236 metaslab_group_alloc_increment(spa_t *spa, uint64_t vdev, void *tag, int flags,
3237     int allocator)
3238 {
3239 	if (!(flags & METASLAB_ASYNC_ALLOC) ||
3240 	    (flags & METASLAB_DONT_THROTTLE))
3241 		return;
3242 
3243 	metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
3244 	if (!mg->mg_class->mc_alloc_throttle_enabled)
3245 		return;
3246 
3247 	(void) zfs_refcount_add(&mg->mg_alloc_queue_depth[allocator], tag);
3248 }
3249 
3250 static void
3251 metaslab_group_increment_qdepth(metaslab_group_t *mg, int allocator)
3252 {
3253 	uint64_t max = mg->mg_max_alloc_queue_depth;
3254 	uint64_t cur = mg->mg_cur_max_alloc_queue_depth[allocator];
3255 	while (cur < max) {
3256 		if (atomic_cas_64(&mg->mg_cur_max_alloc_queue_depth[allocator],
3257 		    cur, cur + 1) == cur) {
3258 			atomic_inc_64(
3259 			    &mg->mg_class->mc_alloc_max_slots[allocator]);
3260 			return;
3261 		}
3262 		cur = mg->mg_cur_max_alloc_queue_depth[allocator];
3263 	}
3264 }
3265 
3266 void
3267 metaslab_group_alloc_decrement(spa_t *spa, uint64_t vdev, void *tag, int flags,
3268     int allocator, boolean_t io_complete)
3269 {
3270 	if (!(flags & METASLAB_ASYNC_ALLOC) ||
3271 	    (flags & METASLAB_DONT_THROTTLE))
3272 		return;
3273 
3274 	metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
3275 	if (!mg->mg_class->mc_alloc_throttle_enabled)
3276 		return;
3277 
3278 	(void) zfs_refcount_remove(&mg->mg_alloc_queue_depth[allocator], tag);
3279 	if (io_complete)
3280 		metaslab_group_increment_qdepth(mg, allocator);
3281 }
3282 
3283 void
3284 metaslab_group_alloc_verify(spa_t *spa, const blkptr_t *bp, void *tag,
3285     int allocator)
3286 {
3287 #ifdef ZFS_DEBUG
3288 	const dva_t *dva = bp->blk_dva;
3289 	int ndvas = BP_GET_NDVAS(bp);
3290 
3291 	for (int d = 0; d < ndvas; d++) {
3292 		uint64_t vdev = DVA_GET_VDEV(&dva[d]);
3293 		metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
3294 		VERIFY(zfs_refcount_not_held(
3295 		    &mg->mg_alloc_queue_depth[allocator], tag));
3296 	}
3297 #endif
3298 }
3299 
3300 static uint64_t
3301 metaslab_block_alloc(metaslab_t *msp, uint64_t size, uint64_t txg)
3302 {
3303 	uint64_t start;
3304 	range_tree_t *rt = msp->ms_allocatable;
3305 	metaslab_class_t *mc = msp->ms_group->mg_class;
3306 
3307 	VERIFY(!msp->ms_condensing);
3308 	VERIFY0(msp->ms_disabled);
3309 
3310 	start = mc->mc_ops->msop_alloc(msp, size);
3311 	if (start != -1ULL) {
3312 		metaslab_group_t *mg = msp->ms_group;
3313 		vdev_t *vd = mg->mg_vd;
3314 
3315 		VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
3316 		VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
3317 		VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
3318 		range_tree_remove(rt, start, size);
3319 		range_tree_clear(msp->ms_trim, start, size);
3320 
3321 		if (range_tree_is_empty(msp->ms_allocating[txg & TXG_MASK]))
3322 			vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
3323 
3324 		range_tree_add(msp->ms_allocating[txg & TXG_MASK], start, size);
3325 
3326 		/* Track the last successful allocation */
3327 		msp->ms_alloc_txg = txg;
3328 		metaslab_verify_space(msp, txg);
3329 	}
3330 
3331 	/*
3332 	 * Now that we've attempted the allocation we need to update the
3333 	 * metaslab's maximum block size since it may have changed.
3334 	 */
3335 	msp->ms_max_size = metaslab_block_maxsize(msp);
3336 	return (start);
3337 }
3338 
3339 /*
3340  * Find the metaslab with the highest weight that is less than what we've
3341  * already tried.  In the common case, this means that we will examine each
3342  * metaslab at most once. Note that concurrent callers could reorder metaslabs
3343  * by activation/passivation once we have dropped the mg_lock. If a metaslab is
3344  * activated by another thread, and we fail to allocate from the metaslab we
3345  * have selected, we may not try the newly-activated metaslab, and instead
3346  * activate another metaslab.  This is not optimal, but generally does not cause
3347  * any problems (a possible exception being if every metaslab is completely full
3348  * except for the the newly-activated metaslab which we fail to examine).
3349  */
3350 static metaslab_t *
3351 find_valid_metaslab(metaslab_group_t *mg, uint64_t activation_weight,
3352     dva_t *dva, int d, boolean_t want_unique, uint64_t asize, int allocator,
3353     zio_alloc_list_t *zal, metaslab_t *search, boolean_t *was_active)
3354 {
3355 	avl_index_t idx;
3356 	avl_tree_t *t = &mg->mg_metaslab_tree;
3357 	metaslab_t *msp = avl_find(t, search, &idx);
3358 	if (msp == NULL)
3359 		msp = avl_nearest(t, idx, AVL_AFTER);
3360 
3361 	for (; msp != NULL; msp = AVL_NEXT(t, msp)) {
3362 		int i;
3363 		if (!metaslab_should_allocate(msp, asize)) {
3364 			metaslab_trace_add(zal, mg, msp, asize, d,
3365 			    TRACE_TOO_SMALL, allocator);
3366 			continue;
3367 		}
3368 
3369 		/*
3370 		 * If the selected metaslab is condensing or disabled,
3371 		 * skip it.
3372 		 */
3373 		if (msp->ms_condensing || msp->ms_disabled > 0)
3374 			continue;
3375 
3376 		*was_active = msp->ms_allocator != -1;
3377 		/*
3378 		 * If we're activating as primary, this is our first allocation
3379 		 * from this disk, so we don't need to check how close we are.
3380 		 * If the metaslab under consideration was already active,
3381 		 * we're getting desperate enough to steal another allocator's
3382 		 * metaslab, so we still don't care about distances.
3383 		 */
3384 		if (activation_weight == METASLAB_WEIGHT_PRIMARY || *was_active)
3385 			break;
3386 
3387 		for (i = 0; i < d; i++) {
3388 			if (want_unique &&
3389 			    !metaslab_is_unique(msp, &dva[i]))
3390 				break;  /* try another metaslab */
3391 		}
3392 		if (i == d)
3393 			break;
3394 	}
3395 
3396 	if (msp != NULL) {
3397 		search->ms_weight = msp->ms_weight;
3398 		search->ms_start = msp->ms_start + 1;
3399 		search->ms_allocator = msp->ms_allocator;
3400 		search->ms_primary = msp->ms_primary;
3401 	}
3402 	return (msp);
3403 }
3404 
3405 /* ARGSUSED */
3406 static uint64_t
3407 metaslab_group_alloc_normal(metaslab_group_t *mg, zio_alloc_list_t *zal,
3408     uint64_t asize, uint64_t txg, boolean_t want_unique, dva_t *dva,
3409     int d, int allocator)
3410 {
3411 	metaslab_t *msp = NULL;
3412 	uint64_t offset = -1ULL;
3413 	uint64_t activation_weight;
3414 
3415 	activation_weight = METASLAB_WEIGHT_PRIMARY;
3416 	for (int i = 0; i < d; i++) {
3417 		if (activation_weight == METASLAB_WEIGHT_PRIMARY &&
3418 		    DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
3419 			activation_weight = METASLAB_WEIGHT_SECONDARY;
3420 		} else if (activation_weight == METASLAB_WEIGHT_SECONDARY &&
3421 		    DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
3422 			activation_weight = METASLAB_WEIGHT_CLAIM;
3423 			break;
3424 		}
3425 	}
3426 
3427 	/*
3428 	 * If we don't have enough metaslabs active to fill the entire array, we
3429 	 * just use the 0th slot.
3430 	 */
3431 	if (mg->mg_ms_ready < mg->mg_allocators * 3)
3432 		allocator = 0;
3433 
3434 	ASSERT3U(mg->mg_vd->vdev_ms_count, >=, 2);
3435 
3436 	metaslab_t *search = kmem_alloc(sizeof (*search), KM_SLEEP);
3437 	search->ms_weight = UINT64_MAX;
3438 	search->ms_start = 0;
3439 	/*
3440 	 * At the end of the metaslab tree are the already-active metaslabs,
3441 	 * first the primaries, then the secondaries. When we resume searching
3442 	 * through the tree, we need to consider ms_allocator and ms_primary so
3443 	 * we start in the location right after where we left off, and don't
3444 	 * accidentally loop forever considering the same metaslabs.
3445 	 */
3446 	search->ms_allocator = -1;
3447 	search->ms_primary = B_TRUE;
3448 	for (;;) {
3449 		boolean_t was_active = B_FALSE;
3450 
3451 		mutex_enter(&mg->mg_lock);
3452 
3453 		if (activation_weight == METASLAB_WEIGHT_PRIMARY &&
3454 		    mg->mg_primaries[allocator] != NULL) {
3455 			msp = mg->mg_primaries[allocator];
3456 			was_active = B_TRUE;
3457 		} else if (activation_weight == METASLAB_WEIGHT_SECONDARY &&
3458 		    mg->mg_secondaries[allocator] != NULL) {
3459 			msp = mg->mg_secondaries[allocator];
3460 			was_active = B_TRUE;
3461 		} else {
3462 			msp = find_valid_metaslab(mg, activation_weight, dva, d,
3463 			    want_unique, asize, allocator, zal, search,
3464 			    &was_active);
3465 		}
3466 
3467 		mutex_exit(&mg->mg_lock);
3468 		if (msp == NULL) {
3469 			kmem_free(search, sizeof (*search));
3470 			return (-1ULL);
3471 		}
3472 
3473 		mutex_enter(&msp->ms_lock);
3474 		/*
3475 		 * Ensure that the metaslab we have selected is still
3476 		 * capable of handling our request. It's possible that
3477 		 * another thread may have changed the weight while we
3478 		 * were blocked on the metaslab lock. We check the
3479 		 * active status first to see if we need to reselect
3480 		 * a new metaslab.
3481 		 */
3482 		if (was_active && !(msp->ms_weight & METASLAB_ACTIVE_MASK)) {
3483 			mutex_exit(&msp->ms_lock);
3484 			continue;
3485 		}
3486 
3487 		/*
3488 		 * If the metaslab is freshly activated for an allocator that
3489 		 * isn't the one we're allocating from, or if it's a primary and
3490 		 * we're seeking a secondary (or vice versa), we go back and
3491 		 * select a new metaslab.
3492 		 */
3493 		if (!was_active && (msp->ms_weight & METASLAB_ACTIVE_MASK) &&
3494 		    (msp->ms_allocator != -1) &&
3495 		    (msp->ms_allocator != allocator || ((activation_weight ==
3496 		    METASLAB_WEIGHT_PRIMARY) != msp->ms_primary))) {
3497 			mutex_exit(&msp->ms_lock);
3498 			continue;
3499 		}
3500 
3501 		if (msp->ms_weight & METASLAB_WEIGHT_CLAIM &&
3502 		    activation_weight != METASLAB_WEIGHT_CLAIM) {
3503 			metaslab_passivate(msp, msp->ms_weight &
3504 			    ~METASLAB_WEIGHT_CLAIM);
3505 			mutex_exit(&msp->ms_lock);
3506 			continue;
3507 		}
3508 
3509 		if (metaslab_activate(msp, allocator, activation_weight) != 0) {
3510 			mutex_exit(&msp->ms_lock);
3511 			continue;
3512 		}
3513 
3514 		msp->ms_selected_txg = txg;
3515 
3516 		/*
3517 		 * Now that we have the lock, recheck to see if we should
3518 		 * continue to use this metaslab for this allocation. The
3519 		 * the metaslab is now loaded so metaslab_should_allocate() can
3520 		 * accurately determine if the allocation attempt should
3521 		 * proceed.
3522 		 */
3523 		if (!metaslab_should_allocate(msp, asize)) {
3524 			/* Passivate this metaslab and select a new one. */
3525 			metaslab_trace_add(zal, mg, msp, asize, d,
3526 			    TRACE_TOO_SMALL, allocator);
3527 			goto next;
3528 		}
3529 
3530 		/*
3531 		 * If this metaslab is currently condensing then pick again as
3532 		 * we can't manipulate this metaslab until it's committed
3533 		 * to disk. If this metaslab is being initialized, we shouldn't
3534 		 * allocate from it since the allocated region might be
3535 		 * overwritten after allocation.
3536 		 */
3537 		if (msp->ms_condensing) {
3538 			metaslab_trace_add(zal, mg, msp, asize, d,
3539 			    TRACE_CONDENSING, allocator);
3540 			metaslab_passivate(msp, msp->ms_weight &
3541 			    ~METASLAB_ACTIVE_MASK);
3542 			mutex_exit(&msp->ms_lock);
3543 			continue;
3544 		} else if (msp->ms_disabled > 0) {
3545 			metaslab_trace_add(zal, mg, msp, asize, d,
3546 			    TRACE_DISABLED, allocator);
3547 			metaslab_passivate(msp, msp->ms_weight &
3548 			    ~METASLAB_ACTIVE_MASK);
3549 			mutex_exit(&msp->ms_lock);
3550 			continue;
3551 		}
3552 
3553 		offset = metaslab_block_alloc(msp, asize, txg);
3554 		metaslab_trace_add(zal, mg, msp, asize, d, offset, allocator);
3555 
3556 		if (offset != -1ULL) {
3557 			/* Proactively passivate the metaslab, if needed */
3558 			metaslab_segment_may_passivate(msp);
3559 			break;
3560 		}
3561 next:
3562 		ASSERT(msp->ms_loaded);
3563 
3564 		/*
3565 		 * We were unable to allocate from this metaslab so determine
3566 		 * a new weight for this metaslab. Now that we have loaded
3567 		 * the metaslab we can provide a better hint to the metaslab
3568 		 * selector.
3569 		 *
3570 		 * For space-based metaslabs, we use the maximum block size.
3571 		 * This information is only available when the metaslab
3572 		 * is loaded and is more accurate than the generic free
3573 		 * space weight that was calculated by metaslab_weight().
3574 		 * This information allows us to quickly compare the maximum
3575 		 * available allocation in the metaslab to the allocation
3576 		 * size being requested.
3577 		 *
3578 		 * For segment-based metaslabs, determine the new weight
3579 		 * based on the highest bucket in the range tree. We
3580 		 * explicitly use the loaded segment weight (i.e. the range
3581 		 * tree histogram) since it contains the space that is
3582 		 * currently available for allocation and is accurate
3583 		 * even within a sync pass.
3584 		 */
3585 		if (WEIGHT_IS_SPACEBASED(msp->ms_weight)) {
3586 			uint64_t weight = metaslab_block_maxsize(msp);
3587 			WEIGHT_SET_SPACEBASED(weight);
3588 			metaslab_passivate(msp, weight);
3589 		} else {
3590 			metaslab_passivate(msp,
3591 			    metaslab_weight_from_range_tree(msp));
3592 		}
3593 
3594 		/*
3595 		 * We have just failed an allocation attempt, check
3596 		 * that metaslab_should_allocate() agrees. Otherwise,
3597 		 * we may end up in an infinite loop retrying the same
3598 		 * metaslab.
3599 		 */
3600 		ASSERT(!metaslab_should_allocate(msp, asize));
3601 
3602 		mutex_exit(&msp->ms_lock);
3603 	}
3604 	mutex_exit(&msp->ms_lock);
3605 	kmem_free(search, sizeof (*search));
3606 	return (offset);
3607 }
3608 
3609 static uint64_t
3610 metaslab_group_alloc(metaslab_group_t *mg, zio_alloc_list_t *zal,
3611     uint64_t asize, uint64_t txg, boolean_t want_unique, dva_t *dva,
3612     int d, int allocator)
3613 {
3614 	uint64_t offset;
3615 	ASSERT(mg->mg_initialized);
3616 
3617 	offset = metaslab_group_alloc_normal(mg, zal, asize, txg, want_unique,
3618 	    dva, d, allocator);
3619 
3620 	mutex_enter(&mg->mg_lock);
3621 	if (offset == -1ULL) {
3622 		mg->mg_failed_allocations++;
3623 		metaslab_trace_add(zal, mg, NULL, asize, d,
3624 		    TRACE_GROUP_FAILURE, allocator);
3625 		if (asize == SPA_GANGBLOCKSIZE) {
3626 			/*
3627 			 * This metaslab group was unable to allocate
3628 			 * the minimum gang block size so it must be out of
3629 			 * space. We must notify the allocation throttle
3630 			 * to start skipping allocation attempts to this
3631 			 * metaslab group until more space becomes available.
3632 			 * Note: this failure cannot be caused by the
3633 			 * allocation throttle since the allocation throttle
3634 			 * is only responsible for skipping devices and
3635 			 * not failing block allocations.
3636 			 */
3637 			mg->mg_no_free_space = B_TRUE;
3638 		}
3639 	}
3640 	mg->mg_allocations++;
3641 	mutex_exit(&mg->mg_lock);
3642 	return (offset);
3643 }
3644 
3645 /*
3646  * Allocate a block for the specified i/o.
3647  */
3648 int
3649 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
3650     dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags,
3651     zio_alloc_list_t *zal, int allocator)
3652 {
3653 	metaslab_group_t *mg, *rotor;
3654 	vdev_t *vd;
3655 	boolean_t try_hard = B_FALSE;
3656 
3657 	ASSERT(!DVA_IS_VALID(&dva[d]));
3658 
3659 	/*
3660 	 * For testing, make some blocks above a certain size be gang blocks.
3661 	 * This will also test spilling from special to normal.
3662 	 */
3663 	if (psize >= metaslab_force_ganging && (ddi_get_lbolt() & 3) == 0) {
3664 		metaslab_trace_add(zal, NULL, NULL, psize, d, TRACE_FORCE_GANG,
3665 		    allocator);
3666 		return (SET_ERROR(ENOSPC));
3667 	}
3668 
3669 	/*
3670 	 * Start at the rotor and loop through all mgs until we find something.
3671 	 * Note that there's no locking on mc_rotor or mc_aliquot because
3672 	 * nothing actually breaks if we miss a few updates -- we just won't
3673 	 * allocate quite as evenly.  It all balances out over time.
3674 	 *
3675 	 * If we are doing ditto or log blocks, try to spread them across
3676 	 * consecutive vdevs.  If we're forced to reuse a vdev before we've
3677 	 * allocated all of our ditto blocks, then try and spread them out on
3678 	 * that vdev as much as possible.  If it turns out to not be possible,
3679 	 * gradually lower our standards until anything becomes acceptable.
3680 	 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
3681 	 * gives us hope of containing our fault domains to something we're
3682 	 * able to reason about.  Otherwise, any two top-level vdev failures
3683 	 * will guarantee the loss of data.  With consecutive allocation,
3684 	 * only two adjacent top-level vdev failures will result in data loss.
3685 	 *
3686 	 * If we are doing gang blocks (hintdva is non-NULL), try to keep
3687 	 * ourselves on the same vdev as our gang block header.  That
3688 	 * way, we can hope for locality in vdev_cache, plus it makes our
3689 	 * fault domains something tractable.
3690 	 */
3691 	if (hintdva) {
3692 		vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
3693 
3694 		/*
3695 		 * It's possible the vdev we're using as the hint no
3696 		 * longer exists or its mg has been closed (e.g. by
3697 		 * device removal).  Consult the rotor when
3698 		 * all else fails.
3699 		 */
3700 		if (vd != NULL && vd->vdev_mg != NULL) {
3701 			mg = vd->vdev_mg;
3702 
3703 			if (flags & METASLAB_HINTBP_AVOID &&
3704 			    mg->mg_next != NULL)
3705 				mg = mg->mg_next;
3706 		} else {
3707 			mg = mc->mc_rotor;
3708 		}
3709 	} else if (d != 0) {
3710 		vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
3711 		mg = vd->vdev_mg->mg_next;
3712 	} else {
3713 		ASSERT(mc->mc_rotor != NULL);
3714 		mg = mc->mc_rotor;
3715 	}
3716 
3717 	/*
3718 	 * If the hint put us into the wrong metaslab class, or into a
3719 	 * metaslab group that has been passivated, just follow the rotor.
3720 	 */
3721 	if (mg->mg_class != mc || mg->mg_activation_count <= 0)
3722 		mg = mc->mc_rotor;
3723 
3724 	rotor = mg;
3725 top:
3726 	do {
3727 		boolean_t allocatable;
3728 
3729 		ASSERT(mg->mg_activation_count == 1);
3730 		vd = mg->mg_vd;
3731 
3732 		/*
3733 		 * Don't allocate from faulted devices.
3734 		 */
3735 		if (try_hard) {
3736 			spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
3737 			allocatable = vdev_allocatable(vd);
3738 			spa_config_exit(spa, SCL_ZIO, FTAG);
3739 		} else {
3740 			allocatable = vdev_allocatable(vd);
3741 		}
3742 
3743 		/*
3744 		 * Determine if the selected metaslab group is eligible
3745 		 * for allocations. If we're ganging then don't allow
3746 		 * this metaslab group to skip allocations since that would
3747 		 * inadvertently return ENOSPC and suspend the pool
3748 		 * even though space is still available.
3749 		 */
3750 		if (allocatable && !GANG_ALLOCATION(flags) && !try_hard) {
3751 			allocatable = metaslab_group_allocatable(mg, rotor,
3752 			    psize, allocator, d);
3753 		}
3754 
3755 		if (!allocatable) {
3756 			metaslab_trace_add(zal, mg, NULL, psize, d,
3757 			    TRACE_NOT_ALLOCATABLE, allocator);
3758 			goto next;
3759 		}
3760 
3761 		ASSERT(mg->mg_initialized);
3762 
3763 		/*
3764 		 * Avoid writing single-copy data to a failing,
3765 		 * non-redundant vdev, unless we've already tried all
3766 		 * other vdevs.
3767 		 */
3768 		if ((vd->vdev_stat.vs_write_errors > 0 ||
3769 		    vd->vdev_state < VDEV_STATE_HEALTHY) &&
3770 		    d == 0 && !try_hard && vd->vdev_children == 0) {
3771 			metaslab_trace_add(zal, mg, NULL, psize, d,
3772 			    TRACE_VDEV_ERROR, allocator);
3773 			goto next;
3774 		}
3775 
3776 		ASSERT(mg->mg_class == mc);
3777 
3778 		uint64_t asize = vdev_psize_to_asize(vd, psize);
3779 		ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
3780 
3781 		/*
3782 		 * If we don't need to try hard, then require that the
3783 		 * block be on an different metaslab from any other DVAs
3784 		 * in this BP (unique=true).  If we are trying hard, then
3785 		 * allow any metaslab to be used (unique=false).
3786 		 */
3787 		uint64_t offset = metaslab_group_alloc(mg, zal, asize, txg,
3788 		    !try_hard, dva, d, allocator);
3789 
3790 		if (offset != -1ULL) {
3791 			/*
3792 			 * If we've just selected this metaslab group,
3793 			 * figure out whether the corresponding vdev is
3794 			 * over- or under-used relative to the pool,
3795 			 * and set an allocation bias to even it out.
3796 			 */
3797 			if (mc->mc_aliquot == 0 && metaslab_bias_enabled) {
3798 				vdev_stat_t *vs = &vd->vdev_stat;
3799 				int64_t vu, cu;
3800 
3801 				vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
3802 				cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
3803 
3804 				/*
3805 				 * Calculate how much more or less we should
3806 				 * try to allocate from this device during
3807 				 * this iteration around the rotor.
3808 				 * For example, if a device is 80% full
3809 				 * and the pool is 20% full then we should
3810 				 * reduce allocations by 60% on this device.
3811 				 *
3812 				 * mg_bias = (20 - 80) * 512K / 100 = -307K
3813 				 *
3814 				 * This reduces allocations by 307K for this
3815 				 * iteration.
3816 				 */
3817 				mg->mg_bias = ((cu - vu) *
3818 				    (int64_t)mg->mg_aliquot) / 100;
3819 			} else if (!metaslab_bias_enabled) {
3820 				mg->mg_bias = 0;
3821 			}
3822 
3823 			if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
3824 			    mg->mg_aliquot + mg->mg_bias) {
3825 				mc->mc_rotor = mg->mg_next;
3826 				mc->mc_aliquot = 0;
3827 			}
3828 
3829 			DVA_SET_VDEV(&dva[d], vd->vdev_id);
3830 			DVA_SET_OFFSET(&dva[d], offset);
3831 			DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
3832 			DVA_SET_ASIZE(&dva[d], asize);
3833 
3834 			return (0);
3835 		}
3836 next:
3837 		mc->mc_rotor = mg->mg_next;
3838 		mc->mc_aliquot = 0;
3839 	} while ((mg = mg->mg_next) != rotor);
3840 
3841 	/*
3842 	 * If we haven't tried hard, do so now.
3843 	 */
3844 	if (!try_hard) {
3845 		try_hard = B_TRUE;
3846 		goto top;
3847 	}
3848 
3849 	bzero(&dva[d], sizeof (dva_t));
3850 
3851 	metaslab_trace_add(zal, rotor, NULL, psize, d, TRACE_ENOSPC, allocator);
3852 	return (SET_ERROR(ENOSPC));
3853 }
3854 
3855 void
3856 metaslab_free_concrete(vdev_t *vd, uint64_t offset, uint64_t asize,
3857     boolean_t checkpoint)
3858 {
3859 	metaslab_t *msp;
3860 	spa_t *spa = vd->vdev_spa;
3861 
3862 	ASSERT(vdev_is_concrete(vd));
3863 	ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
3864 	ASSERT3U(offset >> vd->vdev_ms_shift, <, vd->vdev_ms_count);
3865 
3866 	msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3867 
3868 	VERIFY(!msp->ms_condensing);
3869 	VERIFY3U(offset, >=, msp->ms_start);
3870 	VERIFY3U(offset + asize, <=, msp->ms_start + msp->ms_size);
3871 	VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
3872 	VERIFY0(P2PHASE(asize, 1ULL << vd->vdev_ashift));
3873 
3874 	metaslab_check_free_impl(vd, offset, asize);
3875 
3876 	mutex_enter(&msp->ms_lock);
3877 	if (range_tree_is_empty(msp->ms_freeing) &&
3878 	    range_tree_is_empty(msp->ms_checkpointing)) {
3879 		vdev_dirty(vd, VDD_METASLAB, msp, spa_syncing_txg(spa));
3880 	}
3881 
3882 	if (checkpoint) {
3883 		ASSERT(spa_has_checkpoint(spa));
3884 		range_tree_add(msp->ms_checkpointing, offset, asize);
3885 	} else {
3886 		range_tree_add(msp->ms_freeing, offset, asize);
3887 	}
3888 	mutex_exit(&msp->ms_lock);
3889 }
3890 
3891 /* ARGSUSED */
3892 void
3893 metaslab_free_impl_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset,
3894     uint64_t size, void *arg)
3895 {
3896 	boolean_t *checkpoint = arg;
3897 
3898 	ASSERT3P(checkpoint, !=, NULL);
3899 
3900 	if (vd->vdev_ops->vdev_op_remap != NULL)
3901 		vdev_indirect_mark_obsolete(vd, offset, size);
3902 	else
3903 		metaslab_free_impl(vd, offset, size, *checkpoint);
3904 }
3905 
3906 static void
3907 metaslab_free_impl(vdev_t *vd, uint64_t offset, uint64_t size,
3908     boolean_t checkpoint)
3909 {
3910 	spa_t *spa = vd->vdev_spa;
3911 
3912 	ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
3913 
3914 	if (spa_syncing_txg(spa) > spa_freeze_txg(spa))
3915 		return;
3916 
3917 	if (spa->spa_vdev_removal != NULL &&
3918 	    spa->spa_vdev_removal->svr_vdev_id == vd->vdev_id &&
3919 	    vdev_is_concrete(vd)) {
3920 		/*
3921 		 * Note: we check if the vdev is concrete because when
3922 		 * we complete the removal, we first change the vdev to be
3923 		 * an indirect vdev (in open context), and then (in syncing
3924 		 * context) clear spa_vdev_removal.
3925 		 */
3926 		free_from_removing_vdev(vd, offset, size);
3927 	} else if (vd->vdev_ops->vdev_op_remap != NULL) {
3928 		vdev_indirect_mark_obsolete(vd, offset, size);
3929 		vd->vdev_ops->vdev_op_remap(vd, offset, size,
3930 		    metaslab_free_impl_cb, &checkpoint);
3931 	} else {
3932 		metaslab_free_concrete(vd, offset, size, checkpoint);
3933 	}
3934 }
3935 
3936 typedef struct remap_blkptr_cb_arg {
3937 	blkptr_t *rbca_bp;
3938 	spa_remap_cb_t rbca_cb;
3939 	vdev_t *rbca_remap_vd;
3940 	uint64_t rbca_remap_offset;
3941 	void *rbca_cb_arg;
3942 } remap_blkptr_cb_arg_t;
3943 
3944 void
3945 remap_blkptr_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset,
3946     uint64_t size, void *arg)
3947 {
3948 	remap_blkptr_cb_arg_t *rbca = arg;
3949 	blkptr_t *bp = rbca->rbca_bp;
3950 
3951 	/* We can not remap split blocks. */
3952 	if (size != DVA_GET_ASIZE(&bp->blk_dva[0]))
3953 		return;
3954 	ASSERT0(inner_offset);
3955 
3956 	if (rbca->rbca_cb != NULL) {
3957 		/*
3958 		 * At this point we know that we are not handling split
3959 		 * blocks and we invoke the callback on the previous
3960 		 * vdev which must be indirect.
3961 		 */
3962 		ASSERT3P(rbca->rbca_remap_vd->vdev_ops, ==, &vdev_indirect_ops);
3963 
3964 		rbca->rbca_cb(rbca->rbca_remap_vd->vdev_id,
3965 		    rbca->rbca_remap_offset, size, rbca->rbca_cb_arg);
3966 
3967 		/* set up remap_blkptr_cb_arg for the next call */
3968 		rbca->rbca_remap_vd = vd;
3969 		rbca->rbca_remap_offset = offset;
3970 	}
3971 
3972 	/*
3973 	 * The phys birth time is that of dva[0].  This ensures that we know
3974 	 * when each dva was written, so that resilver can determine which
3975 	 * blocks need to be scrubbed (i.e. those written during the time
3976 	 * the vdev was offline).  It also ensures that the key used in
3977 	 * the ARC hash table is unique (i.e. dva[0] + phys_birth).  If
3978 	 * we didn't change the phys_birth, a lookup in the ARC for a
3979 	 * remapped BP could find the data that was previously stored at
3980 	 * this vdev + offset.
3981 	 */
3982 	vdev_t *oldvd = vdev_lookup_top(vd->vdev_spa,
3983 	    DVA_GET_VDEV(&bp->blk_dva[0]));
3984 	vdev_indirect_births_t *vib = oldvd->vdev_indirect_births;
3985 	bp->blk_phys_birth = vdev_indirect_births_physbirth(vib,
3986 	    DVA_GET_OFFSET(&bp->blk_dva[0]), DVA_GET_ASIZE(&bp->blk_dva[0]));
3987 
3988 	DVA_SET_VDEV(&bp->blk_dva[0], vd->vdev_id);
3989 	DVA_SET_OFFSET(&bp->blk_dva[0], offset);
3990 }
3991 
3992 /*
3993  * If the block pointer contains any indirect DVAs, modify them to refer to
3994  * concrete DVAs.  Note that this will sometimes not be possible, leaving
3995  * the indirect DVA in place.  This happens if the indirect DVA spans multiple
3996  * segments in the mapping (i.e. it is a "split block").
3997  *
3998  * If the BP was remapped, calls the callback on the original dva (note the
3999  * callback can be called multiple times if the original indirect DVA refers
4000  * to another indirect DVA, etc).
4001  *
4002  * Returns TRUE if the BP was remapped.
4003  */
4004 boolean_t
4005 spa_remap_blkptr(spa_t *spa, blkptr_t *bp, spa_remap_cb_t callback, void *arg)
4006 {
4007 	remap_blkptr_cb_arg_t rbca;
4008 
4009 	if (!zfs_remap_blkptr_enable)
4010 		return (B_FALSE);
4011 
4012 	if (!spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS))
4013 		return (B_FALSE);
4014 
4015 	/*
4016 	 * Dedup BP's can not be remapped, because ddt_phys_select() depends
4017 	 * on DVA[0] being the same in the BP as in the DDT (dedup table).
4018 	 */
4019 	if (BP_GET_DEDUP(bp))
4020 		return (B_FALSE);
4021 
4022 	/*
4023 	 * Gang blocks can not be remapped, because
4024 	 * zio_checksum_gang_verifier() depends on the DVA[0] that's in
4025 	 * the BP used to read the gang block header (GBH) being the same
4026 	 * as the DVA[0] that we allocated for the GBH.
4027 	 */
4028 	if (BP_IS_GANG(bp))
4029 		return (B_FALSE);
4030 
4031 	/*
4032 	 * Embedded BP's have no DVA to remap.
4033 	 */
4034 	if (BP_GET_NDVAS(bp) < 1)
4035 		return (B_FALSE);
4036 
4037 	/*
4038 	 * Note: we only remap dva[0].  If we remapped other dvas, we
4039 	 * would no longer know what their phys birth txg is.
4040 	 */
4041 	dva_t *dva = &bp->blk_dva[0];
4042 
4043 	uint64_t offset = DVA_GET_OFFSET(dva);
4044 	uint64_t size = DVA_GET_ASIZE(dva);
4045 	vdev_t *vd = vdev_lookup_top(spa, DVA_GET_VDEV(dva));
4046 
4047 	if (vd->vdev_ops->vdev_op_remap == NULL)
4048 		return (B_FALSE);
4049 
4050 	rbca.rbca_bp = bp;
4051 	rbca.rbca_cb = callback;
4052 	rbca.rbca_remap_vd = vd;
4053 	rbca.rbca_remap_offset = offset;
4054 	rbca.rbca_cb_arg = arg;
4055 
4056 	/*
4057 	 * remap_blkptr_cb() will be called in order for each level of
4058 	 * indirection, until a concrete vdev is reached or a split block is
4059 	 * encountered. old_vd and old_offset are updated within the callback
4060 	 * as we go from the one indirect vdev to the next one (either concrete
4061 	 * or indirect again) in that order.
4062 	 */
4063 	vd->vdev_ops->vdev_op_remap(vd, offset, size, remap_blkptr_cb, &rbca);
4064 
4065 	/* Check if the DVA wasn't remapped because it is a split block */
4066 	if (DVA_GET_VDEV(&rbca.rbca_bp->blk_dva[0]) == vd->vdev_id)
4067 		return (B_FALSE);
4068 
4069 	return (B_TRUE);
4070 }
4071 
4072 /*
4073  * Undo the allocation of a DVA which happened in the given transaction group.
4074  */
4075 void
4076 metaslab_unalloc_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
4077 {
4078 	metaslab_t *msp;
4079 	vdev_t *vd;
4080 	uint64_t vdev = DVA_GET_VDEV(dva);
4081 	uint64_t offset = DVA_GET_OFFSET(dva);
4082 	uint64_t size = DVA_GET_ASIZE(dva);
4083 
4084 	ASSERT(DVA_IS_VALID(dva));
4085 	ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
4086 
4087 	if (txg > spa_freeze_txg(spa))
4088 		return;
4089 
4090 	if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
4091 	    (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
4092 		cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
4093 		    (u_longlong_t)vdev, (u_longlong_t)offset);
4094 		ASSERT(0);
4095 		return;
4096 	}
4097 
4098 	ASSERT(!vd->vdev_removing);
4099 	ASSERT(vdev_is_concrete(vd));
4100 	ASSERT0(vd->vdev_indirect_config.vic_mapping_object);
4101 	ASSERT3P(vd->vdev_indirect_mapping, ==, NULL);
4102 
4103 	if (DVA_GET_GANG(dva))
4104 		size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
4105 
4106 	msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
4107 
4108 	mutex_enter(&msp->ms_lock);
4109 	range_tree_remove(msp->ms_allocating[txg & TXG_MASK],
4110 	    offset, size);
4111 
4112 	VERIFY(!msp->ms_condensing);
4113 	VERIFY3U(offset, >=, msp->ms_start);
4114 	VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size);
4115 	VERIFY3U(range_tree_space(msp->ms_allocatable) + size, <=,
4116 	    msp->ms_size);
4117 	VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
4118 	VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
4119 	range_tree_add(msp->ms_allocatable, offset, size);
4120 	mutex_exit(&msp->ms_lock);
4121 }
4122 
4123 /*
4124  * Free the block represented by the given DVA.
4125  */
4126 void
4127 metaslab_free_dva(spa_t *spa, const dva_t *dva, boolean_t checkpoint)
4128 {
4129 	uint64_t vdev = DVA_GET_VDEV(dva);
4130 	uint64_t offset = DVA_GET_OFFSET(dva);
4131 	uint64_t size = DVA_GET_ASIZE(dva);
4132 	vdev_t *vd = vdev_lookup_top(spa, vdev);
4133 
4134 	ASSERT(DVA_IS_VALID(dva));
4135 	ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
4136 
4137 	if (DVA_GET_GANG(dva)) {
4138 		size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
4139 	}
4140 
4141 	metaslab_free_impl(vd, offset, size, checkpoint);
4142 }
4143 
4144 /*
4145  * Reserve some allocation slots. The reservation system must be called
4146  * before we call into the allocator. If there aren't any available slots
4147  * then the I/O will be throttled until an I/O completes and its slots are
4148  * freed up. The function returns true if it was successful in placing
4149  * the reservation.
4150  */
4151 boolean_t
4152 metaslab_class_throttle_reserve(metaslab_class_t *mc, int slots, int allocator,
4153     zio_t *zio, int flags)
4154 {
4155 	uint64_t available_slots = 0;
4156 	boolean_t slot_reserved = B_FALSE;
4157 	uint64_t max = mc->mc_alloc_max_slots[allocator];
4158 
4159 	ASSERT(mc->mc_alloc_throttle_enabled);
4160 	mutex_enter(&mc->mc_lock);
4161 
4162 	uint64_t reserved_slots =
4163 	    zfs_refcount_count(&mc->mc_alloc_slots[allocator]);
4164 	if (reserved_slots < max)
4165 		available_slots = max - reserved_slots;
4166 
4167 	if (slots <= available_slots || GANG_ALLOCATION(flags) ||
4168 	    flags & METASLAB_MUST_RESERVE) {
4169 		/*
4170 		 * We reserve the slots individually so that we can unreserve
4171 		 * them individually when an I/O completes.
4172 		 */
4173 		for (int d = 0; d < slots; d++) {
4174 			reserved_slots =
4175 			    zfs_refcount_add(&mc->mc_alloc_slots[allocator],
4176 			    zio);
4177 		}
4178 		zio->io_flags |= ZIO_FLAG_IO_ALLOCATING;
4179 		slot_reserved = B_TRUE;
4180 	}
4181 
4182 	mutex_exit(&mc->mc_lock);
4183 	return (slot_reserved);
4184 }
4185 
4186 void
4187 metaslab_class_throttle_unreserve(metaslab_class_t *mc, int slots,
4188     int allocator, zio_t *zio)
4189 {
4190 	ASSERT(mc->mc_alloc_throttle_enabled);
4191 	mutex_enter(&mc->mc_lock);
4192 	for (int d = 0; d < slots; d++) {
4193 		(void) zfs_refcount_remove(&mc->mc_alloc_slots[allocator],
4194 		    zio);
4195 	}
4196 	mutex_exit(&mc->mc_lock);
4197 }
4198 
4199 static int
4200 metaslab_claim_concrete(vdev_t *vd, uint64_t offset, uint64_t size,
4201     uint64_t txg)
4202 {
4203 	metaslab_t *msp;
4204 	spa_t *spa = vd->vdev_spa;
4205 	int error = 0;
4206 
4207 	if (offset >> vd->vdev_ms_shift >= vd->vdev_ms_count)
4208 		return (ENXIO);
4209 
4210 	ASSERT3P(vd->vdev_ms, !=, NULL);
4211 	msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
4212 
4213 	mutex_enter(&msp->ms_lock);
4214 
4215 	if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded)
4216 		error = metaslab_activate(msp, 0, METASLAB_WEIGHT_CLAIM);
4217 	/*
4218 	 * No need to fail in that case; someone else has activated the
4219 	 * metaslab, but that doesn't preclude us from using it.
4220 	 */
4221 	if (error == EBUSY)
4222 		error = 0;
4223 
4224 	if (error == 0 &&
4225 	    !range_tree_contains(msp->ms_allocatable, offset, size))
4226 		error = SET_ERROR(ENOENT);
4227 
4228 	if (error || txg == 0) {	/* txg == 0 indicates dry run */
4229 		mutex_exit(&msp->ms_lock);
4230 		return (error);
4231 	}
4232 
4233 	VERIFY(!msp->ms_condensing);
4234 	VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
4235 	VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
4236 	VERIFY3U(range_tree_space(msp->ms_allocatable) - size, <=,
4237 	    msp->ms_size);
4238 	range_tree_remove(msp->ms_allocatable, offset, size);
4239 	range_tree_clear(msp->ms_trim, offset, size);
4240 
4241 	if (spa_writeable(spa)) {	/* don't dirty if we're zdb(1M) */
4242 		if (range_tree_is_empty(msp->ms_allocating[txg & TXG_MASK]))
4243 			vdev_dirty(vd, VDD_METASLAB, msp, txg);
4244 		range_tree_add(msp->ms_allocating[txg & TXG_MASK],
4245 		    offset, size);
4246 	}
4247 
4248 	mutex_exit(&msp->ms_lock);
4249 
4250 	return (0);
4251 }
4252 
4253 typedef struct metaslab_claim_cb_arg_t {
4254 	uint64_t	mcca_txg;
4255 	int		mcca_error;
4256 } metaslab_claim_cb_arg_t;
4257 
4258 /* ARGSUSED */
4259 static void
4260 metaslab_claim_impl_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset,
4261     uint64_t size, void *arg)
4262 {
4263 	metaslab_claim_cb_arg_t *mcca_arg = arg;
4264 
4265 	if (mcca_arg->mcca_error == 0) {
4266 		mcca_arg->mcca_error = metaslab_claim_concrete(vd, offset,
4267 		    size, mcca_arg->mcca_txg);
4268 	}
4269 }
4270 
4271 int
4272 metaslab_claim_impl(vdev_t *vd, uint64_t offset, uint64_t size, uint64_t txg)
4273 {
4274 	if (vd->vdev_ops->vdev_op_remap != NULL) {
4275 		metaslab_claim_cb_arg_t arg;
4276 
4277 		/*
4278 		 * Only zdb(1M) can claim on indirect vdevs.  This is used
4279 		 * to detect leaks of mapped space (that are not accounted
4280 		 * for in the obsolete counts, spacemap, or bpobj).
4281 		 */
4282 		ASSERT(!spa_writeable(vd->vdev_spa));
4283 		arg.mcca_error = 0;
4284 		arg.mcca_txg = txg;
4285 
4286 		vd->vdev_ops->vdev_op_remap(vd, offset, size,
4287 		    metaslab_claim_impl_cb, &arg);
4288 
4289 		if (arg.mcca_error == 0) {
4290 			arg.mcca_error = metaslab_claim_concrete(vd,
4291 			    offset, size, txg);
4292 		}
4293 		return (arg.mcca_error);
4294 	} else {
4295 		return (metaslab_claim_concrete(vd, offset, size, txg));
4296 	}
4297 }
4298 
4299 /*
4300  * Intent log support: upon opening the pool after a crash, notify the SPA
4301  * of blocks that the intent log has allocated for immediate write, but
4302  * which are still considered free by the SPA because the last transaction
4303  * group didn't commit yet.
4304  */
4305 static int
4306 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
4307 {
4308 	uint64_t vdev = DVA_GET_VDEV(dva);
4309 	uint64_t offset = DVA_GET_OFFSET(dva);
4310 	uint64_t size = DVA_GET_ASIZE(dva);
4311 	vdev_t *vd;
4312 
4313 	if ((vd = vdev_lookup_top(spa, vdev)) == NULL) {
4314 		return (SET_ERROR(ENXIO));
4315 	}
4316 
4317 	ASSERT(DVA_IS_VALID(dva));
4318 
4319 	if (DVA_GET_GANG(dva))
4320 		size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
4321 
4322 	return (metaslab_claim_impl(vd, offset, size, txg));
4323 }
4324 
4325 int
4326 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
4327     int ndvas, uint64_t txg, blkptr_t *hintbp, int flags,
4328     zio_alloc_list_t *zal, zio_t *zio, int allocator)
4329 {
4330 	dva_t *dva = bp->blk_dva;
4331 	dva_t *hintdva = (hintbp != NULL) ? hintbp->blk_dva : NULL;
4332 	int error = 0;
4333 
4334 	ASSERT(bp->blk_birth == 0);
4335 	ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
4336 
4337 	spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
4338 
4339 	if (mc->mc_rotor == NULL) {	/* no vdevs in this class */
4340 		spa_config_exit(spa, SCL_ALLOC, FTAG);
4341 		return (SET_ERROR(ENOSPC));
4342 	}
4343 
4344 	ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
4345 	ASSERT(BP_GET_NDVAS(bp) == 0);
4346 	ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
4347 	ASSERT3P(zal, !=, NULL);
4348 
4349 	for (int d = 0; d < ndvas; d++) {
4350 		error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
4351 		    txg, flags, zal, allocator);
4352 		if (error != 0) {
4353 			for (d--; d >= 0; d--) {
4354 				metaslab_unalloc_dva(spa, &dva[d], txg);
4355 				metaslab_group_alloc_decrement(spa,
4356 				    DVA_GET_VDEV(&dva[d]), zio, flags,
4357 				    allocator, B_FALSE);
4358 				bzero(&dva[d], sizeof (dva_t));
4359 			}
4360 			spa_config_exit(spa, SCL_ALLOC, FTAG);
4361 			return (error);
4362 		} else {
4363 			/*
4364 			 * Update the metaslab group's queue depth
4365 			 * based on the newly allocated dva.
4366 			 */
4367 			metaslab_group_alloc_increment(spa,
4368 			    DVA_GET_VDEV(&dva[d]), zio, flags, allocator);
4369 		}
4370 
4371 	}
4372 	ASSERT(error == 0);
4373 	ASSERT(BP_GET_NDVAS(bp) == ndvas);
4374 
4375 	spa_config_exit(spa, SCL_ALLOC, FTAG);
4376 
4377 	BP_SET_BIRTH(bp, txg, txg);
4378 
4379 	return (0);
4380 }
4381 
4382 void
4383 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
4384 {
4385 	const dva_t *dva = bp->blk_dva;
4386 	int ndvas = BP_GET_NDVAS(bp);
4387 
4388 	ASSERT(!BP_IS_HOLE(bp));
4389 	ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
4390 
4391 	/*
4392 	 * If we have a checkpoint for the pool we need to make sure that
4393 	 * the blocks that we free that are part of the checkpoint won't be
4394 	 * reused until the checkpoint is discarded or we revert to it.
4395 	 *
4396 	 * The checkpoint flag is passed down the metaslab_free code path
4397 	 * and is set whenever we want to add a block to the checkpoint's
4398 	 * accounting. That is, we "checkpoint" blocks that existed at the
4399 	 * time the checkpoint was created and are therefore referenced by
4400 	 * the checkpointed uberblock.
4401 	 *
4402 	 * Note that, we don't checkpoint any blocks if the current
4403 	 * syncing txg <= spa_checkpoint_txg. We want these frees to sync
4404 	 * normally as they will be referenced by the checkpointed uberblock.
4405 	 */
4406 	boolean_t checkpoint = B_FALSE;
4407 	if (bp->blk_birth <= spa->spa_checkpoint_txg &&
4408 	    spa_syncing_txg(spa) > spa->spa_checkpoint_txg) {
4409 		/*
4410 		 * At this point, if the block is part of the checkpoint
4411 		 * there is no way it was created in the current txg.
4412 		 */
4413 		ASSERT(!now);
4414 		ASSERT3U(spa_syncing_txg(spa), ==, txg);
4415 		checkpoint = B_TRUE;
4416 	}
4417 
4418 	spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
4419 
4420 	for (int d = 0; d < ndvas; d++) {
4421 		if (now) {
4422 			metaslab_unalloc_dva(spa, &dva[d], txg);
4423 		} else {
4424 			ASSERT3U(txg, ==, spa_syncing_txg(spa));
4425 			metaslab_free_dva(spa, &dva[d], checkpoint);
4426 		}
4427 	}
4428 
4429 	spa_config_exit(spa, SCL_FREE, FTAG);
4430 }
4431 
4432 int
4433 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
4434 {
4435 	const dva_t *dva = bp->blk_dva;
4436 	int ndvas = BP_GET_NDVAS(bp);
4437 	int error = 0;
4438 
4439 	ASSERT(!BP_IS_HOLE(bp));
4440 
4441 	if (txg != 0) {
4442 		/*
4443 		 * First do a dry run to make sure all DVAs are claimable,
4444 		 * so we don't have to unwind from partial failures below.
4445 		 */
4446 		if ((error = metaslab_claim(spa, bp, 0)) != 0)
4447 			return (error);
4448 	}
4449 
4450 	spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
4451 
4452 	for (int d = 0; d < ndvas; d++) {
4453 		error = metaslab_claim_dva(spa, &dva[d], txg);
4454 		if (error != 0)
4455 			break;
4456 	}
4457 
4458 	spa_config_exit(spa, SCL_ALLOC, FTAG);
4459 
4460 	ASSERT(error == 0 || txg == 0);
4461 
4462 	return (error);
4463 }
4464 
4465 /* ARGSUSED */
4466 static void
4467 metaslab_check_free_impl_cb(uint64_t inner, vdev_t *vd, uint64_t offset,
4468     uint64_t size, void *arg)
4469 {
4470 	if (vd->vdev_ops == &vdev_indirect_ops)
4471 		return;
4472 
4473 	metaslab_check_free_impl(vd, offset, size);
4474 }
4475 
4476 static void
4477 metaslab_check_free_impl(vdev_t *vd, uint64_t offset, uint64_t size)
4478 {
4479 	metaslab_t *msp;
4480 	spa_t *spa = vd->vdev_spa;
4481 
4482 	if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
4483 		return;
4484 
4485 	if (vd->vdev_ops->vdev_op_remap != NULL) {
4486 		vd->vdev_ops->vdev_op_remap(vd, offset, size,
4487 		    metaslab_check_free_impl_cb, NULL);
4488 		return;
4489 	}
4490 
4491 	ASSERT(vdev_is_concrete(vd));
4492 	ASSERT3U(offset >> vd->vdev_ms_shift, <, vd->vdev_ms_count);
4493 	ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
4494 
4495 	msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
4496 
4497 	mutex_enter(&msp->ms_lock);
4498 	if (msp->ms_loaded) {
4499 		range_tree_verify_not_present(msp->ms_allocatable,
4500 		    offset, size);
4501 	}
4502 
4503 	range_tree_verify_not_present(msp->ms_trim, offset, size);
4504 	range_tree_verify_not_present(msp->ms_freeing, offset, size);
4505 	range_tree_verify_not_present(msp->ms_checkpointing, offset, size);
4506 	range_tree_verify_not_present(msp->ms_freed, offset, size);
4507 	for (int j = 0; j < TXG_DEFER_SIZE; j++)
4508 		range_tree_verify_not_present(msp->ms_defer[j], offset, size);
4509 	mutex_exit(&msp->ms_lock);
4510 }
4511 
4512 void
4513 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
4514 {
4515 	if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
4516 		return;
4517 
4518 	spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
4519 	for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
4520 		uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
4521 		vdev_t *vd = vdev_lookup_top(spa, vdev);
4522 		uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
4523 		uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
4524 
4525 		if (DVA_GET_GANG(&bp->blk_dva[i]))
4526 			size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
4527 
4528 		ASSERT3P(vd, !=, NULL);
4529 
4530 		metaslab_check_free_impl(vd, offset, size);
4531 	}
4532 	spa_config_exit(spa, SCL_VDEV, FTAG);
4533 }
4534 
4535 static void
4536 metaslab_group_disable_wait(metaslab_group_t *mg)
4537 {
4538 	ASSERT(MUTEX_HELD(&mg->mg_ms_disabled_lock));
4539 	while (mg->mg_disabled_updating) {
4540 		cv_wait(&mg->mg_ms_disabled_cv, &mg->mg_ms_disabled_lock);
4541 	}
4542 }
4543 
4544 static void
4545 metaslab_group_disabled_increment(metaslab_group_t *mg)
4546 {
4547 	ASSERT(MUTEX_HELD(&mg->mg_ms_disabled_lock));
4548 	ASSERT(mg->mg_disabled_updating);
4549 
4550 	while (mg->mg_ms_disabled >= max_disabled_ms) {
4551 		cv_wait(&mg->mg_ms_disabled_cv, &mg->mg_ms_disabled_lock);
4552 	}
4553 	mg->mg_ms_disabled++;
4554 	ASSERT3U(mg->mg_ms_disabled, <=, max_disabled_ms);
4555 }
4556 
4557 /*
4558  * Mark the metaslab as disabled to prevent any allocations on this metaslab.
4559  * We must also track how many metaslabs are currently disabled within a
4560  * metaslab group and limit them to prevent allocation failures from
4561  * occurring because all metaslabs are disabled.
4562  */
4563 void
4564 metaslab_disable(metaslab_t *msp)
4565 {
4566 	ASSERT(!MUTEX_HELD(&msp->ms_lock));
4567 	metaslab_group_t *mg = msp->ms_group;
4568 
4569 	mutex_enter(&mg->mg_ms_disabled_lock);
4570 
4571 	/*
4572 	 * To keep an accurate count of how many threads have disabled
4573 	 * a specific metaslab group, we only allow one thread to mark
4574 	 * the metaslab group at a time. This ensures that the value of
4575 	 * ms_disabled will be accurate when we decide to mark a metaslab
4576 	 * group as disabled. To do this we force all other threads
4577 	 * to wait till the metaslab's mg_disabled_updating flag is no
4578 	 * longer set.
4579 	 */
4580 	metaslab_group_disable_wait(mg);
4581 	mg->mg_disabled_updating = B_TRUE;
4582 	if (msp->ms_disabled == 0) {
4583 		metaslab_group_disabled_increment(mg);
4584 	}
4585 	mutex_enter(&msp->ms_lock);
4586 	msp->ms_disabled++;
4587 	mutex_exit(&msp->ms_lock);
4588 
4589 	mg->mg_disabled_updating = B_FALSE;
4590 	cv_broadcast(&mg->mg_ms_disabled_cv);
4591 	mutex_exit(&mg->mg_ms_disabled_lock);
4592 }
4593 
4594 void
4595 metaslab_enable(metaslab_t *msp, boolean_t sync)
4596 {
4597 	metaslab_group_t *mg = msp->ms_group;
4598 	spa_t *spa = mg->mg_vd->vdev_spa;
4599 
4600 	/*
4601 	 * Wait for the outstanding IO to be synced to prevent newly
4602 	 * allocated blocks from being overwritten.  This used by
4603 	 * initialize and TRIM which are modifying unallocated space.
4604 	 */
4605 	if (sync)
4606 		txg_wait_synced(spa_get_dsl(spa), 0);
4607 
4608 	mutex_enter(&mg->mg_ms_disabled_lock);
4609 	mutex_enter(&msp->ms_lock);
4610 	if (--msp->ms_disabled == 0) {
4611 		mg->mg_ms_disabled--;
4612 		cv_broadcast(&mg->mg_ms_disabled_cv);
4613 	}
4614 	mutex_exit(&msp->ms_lock);
4615 	mutex_exit(&mg->mg_ms_disabled_lock);
4616 }
4617